Directions Brain

log

64 min read

Log

Chronological, append-only operations log. One H2 per entry.

Entry format (greppable):

## [YYYY-MM-DD] <op> | <title>

Where <op> is one of: ingest, query, lint, refactor. Body (under each H2) is freeform: brief notes, pages touched, links to relevant [[wiki-pages]].

[2026-05-09] ingest | meeting: 2026-05-06-brain-os-strategy-brainstorm

[2026-05-09] refactor | introduce wiki/action-items.md

  • reason: User wants action items monitored across meetings, not just buried inside each meeting page.
  • touched: CLAUDE.md (Section A invariants 4+5 clarified to exempt operational root files; Section B directory map and narrative; Section F new step 8 for action-items maintenance, log append renumbered to step 9; Section I lint checks added), wiki/action-items.md (created), README.md (layout tree + new “Track action items” subsection).
  • notes: Single rolling page architecture chosen over per-item files (lighter ceremony, easier scan). Backfilled 5 open items from 2026-05-06-brain-os-strategy-brainstorm. wiki/action-items.md is the third operational root file alongside index.md and log.md — it has no frontmatter and is excluded from index.md.

[2026-05-09] refactor | document URL-fetch acquisition workflow for articles

  • reason: User asked how to add online sources for competitor research (starting with glean). Schema previously assumed raw/articles/ files were user-dropped only; needed an LLM-driven acquisition path.
  • touched: CLAUDE.md (Section A invariant 1 carve-out for new article files; Section G expanded into G.0 acquisition + G.1 ingest steps; provenance block format defined; log entry now includes url: line for traceability), README.md (“Add an article” section rewritten to lead with the URL workflow and keep manual drop as fallback).
  • notes: Default acquisition uses axcli chrome navigate/read/close so the user’s authenticated browser handles paywalls and logins; raw markdown is captured with no LLM summarization in the middle. The ONLY content the LLM may write into raw/ is a 4-line provenance block (Source URL + Captured date + Captured by); files remain immutable thereafter.

[2026-05-27] ingest | meeting: 2026-05-27-directions-vertical-pivot-and-prediction-markets

[2026-05-27] refactor | rename hari → ari-leshno (identity collision from prior ingest)

  • reason: ASR mis-rendering surfaced when ingesting raw/meetings/notetaker/2026-05-19-saar-leshno-15-00.json. The 2026-05-27 ingest above had captured the same person as a hari stub from spoken Hebrew “הארי” (the name “ארי” with a definite article). Per Section J dedup (alias match against the new transcript’s “Ari Leshno”), merge into a single page.
  • touched: wiki/people/ari-leshno.md (created — combines original hari.md content with 2026-05-19 enrichment), wiki/people/hari.md (deleted), wiki/meetings/2026-05-27-directions-vertical-pivot-and-prediction-markets.md (3 wikilinks updated), wiki/topics/vertical-use-case-led-brain.md (2 wikilinks updated), wiki/people/saar-arbel.md (2 wikilinks updated), wiki/index.md (People entry renamed)
  • notes: New aliases on [[ari-leshno]] retain “Hari” / “הארי” so future ASR mis-renderings match by alias per Section J. All original Hari content (clinic-operations anchor, vertical-Brain anecdote) is preserved on the new page; 2026-05-19 details (Sheba role, glaucoma pitch, MVP, sizing, FDA precedent) are added.

[2026-05-27] ingest | meeting: 2026-05-19-saar-leshno-glaucoma-data-tracking

[2026-05-30] ingest | source: 2026-05-01-entree-capital-enterprise-ai-spend-map

  • source: raw/articles/vc-articles/2026-05-01-entree-capital-enterprise-ai-spend-map.pdf (PDF, 13 slides + Sources & footnotes; supplied directly by user — not web-captured. Originally named Enterprise AI Spend Map.pdf; renamed to convention later the same day — see refactor entry below.)
  • url: (not applicable — PDF supplied directly; no provenance block in raw)
  • new: 2026-05-01-entree-capital-enterprise-ai-spend-map, entree-capital, openai, anthropic, vertical-ai-tam, slm-default-and-fine-tuning, foundation-labs-monetization-shift
  • updated: vertical-use-case-led-brain (source_count 2→3; new Evidence row citing Bessemer +400% YoY, 752B/2029, Anthropic $70B/2028; Related extended with foundation-labs-monetization-shift), clinical-data-portability (source_count 1→2; new Evidence row on healthcare as 2026-27 vertical breakout; Related extended with vertical-ai-tam), cursor (source_count 2→3; Mentioned-in row added), glean (source_count 1→2; Mentioned-in row added), sierra (source_count 1→2; Mentioned-in row added), wiki/index.md (3 new Companies, 1 new Source, 3 new Topics)
  • notes: First VC-published source in the wiki. Established new tag pattern investor-thesis to flag VC-published reports as advocacy-shaped narrative rather than neutral analysis (applied to source page + entree-capital + the three new topic pages where relevant). Date 2026-05-01 chosen over today (2026-05-30) per user — PDF declares “May 2026” only, no specific day. Entity-stub scope: Section F step 4 (“create stub for each referenced entity”) was applied selectively — created stubs only for openai and anthropic because they are the load-bearing subjects of slide 5’s monetization-divergence claim. Other entities mentioned only as exemplars (Microsoft / Salesforce / ServiceNow / Google / Harvey / Bessemer / Menlo / a16z / Gartner) were not stubbed; future sources can promote them when needed. No action items: market-overview source, not a meeting; no follow-ups generated. Quality flag for entree-capital: stated thesis aligns suspiciously well with the vertical-use-case-led-brain direction the team had already chosen — read with the same skepticism applied to any VC’s portfolio-aligned market map.

[2026-05-30] refactor | rename raw PDF to YYYY-MM-DD-publisher-slug convention

  • reason: User requested the raw filename match the YYYY-MM-DD-publisher-slug convention (Section C). Original drop was Enterprise AI Spend Map.pdf with no date or publisher prefix.
  • touched: raw/articles/Enterprise AI Spend Map.pdf → raw/articles/vc-articles/2026-05-01-entree-capital-enterprise-ai-spend-map.pdf (renamed); wiki/sources/2026-05-01-entree-capital-enterprise-ai-spend-map.md (Source: line + Note paragraph updated); wiki/log.md (preceding ingest entry’s source: line annotated with rename history).
  • notes: Section A invariant 1 was overridden for this rename — the invariant says NEVER modify, delete, or rename existing files under raw/. User explicitly asked for the rename, treated as a one-time exception. If this should be the new default, CLAUDE.md Section A invariant 1 needs to be relaxed to allow agent-driven rename-to-convention (probably with a tighter rule: “may rename to match Section C convention; may not modify content”). Flagging here so future-Claude doesn’t quietly normalize the rule.

[2026-05-30] ingest | source: alon-huri LinkedIn — pilot ingest (3 posts) + latest-15 sweep (15 posts total)

  • source: raw/articles/linkedin-posts/2026-05-29-alonhuri-linkedin-creator-arbitrage.md, raw/articles/linkedin-posts/2026-05-27-alonhuri-linkedin-soldiers-ptsd-update.md, raw/articles/linkedin-posts/2026-05-22-alonhuri-linkedin-reddit-personal-brand.md, raw/articles/linkedin-posts/2026-05-16-alonhuri-linkedin-ai-native-growth-hacking.md, raw/articles/linkedin-posts/2026-05-09-alonhuri-linkedin-ai-native-company.md, raw/articles/linkedin-posts/2026-03-06-alonhuri-linkedin-saas-is-dead-ai-native-agency.md, raw/articles/linkedin-posts/2026-02-27-alonhuri-linkedin-reddit-experiment-results.md, raw/articles/linkedin-posts/2026-02-20-alonhuri-linkedin-ads-in-ai-chats.md, raw/articles/linkedin-posts/2026-02-06-alonhuri-linkedin-bootstrap-vs-astrap.md, raw/articles/linkedin-posts/2026-01-30-alonhuri-linkedin-full-stack-ai.md, raw/articles/linkedin-posts/2026-01-23-alonhuri-linkedin-vibe-coding-2.md, raw/articles/linkedin-posts/2026-01-16-alonhuri-linkedin-soldiers-ptsd-original.md, raw/articles/linkedin-posts/2026-01-09-alonhuri-linkedin-vibe-acting.md, raw/articles/linkedin-posts/2026-01-02-alonhuri-linkedin-100-ideation-hacks.md, raw/articles/linkedin-posts/2025-12-26-alonhuri-linkedin-apple-openai-merger.md. 15 raw files, all Hebrew, captured via axcli chrome from user’s authenticated session. Posts 1, 4, 5 captured directly from the user-supplied LinkedIn URLs in the pilot phase; posts 2, 3, 16–25 captured from the LinkedIn activity feed (linkedin.com/in/alonhuri/recent-activity/all/) using axcli chrome click to trigger lazy-loading and sed to slice the resulting page-capture by feed-post index.
  • url: per-post LinkedIn share URLs available for pilot posts (in their respective source-page frontmatter); for the activity-feed sweep, individual share URLs were not exposed in the accessibility tree — provenance points to the feed URL with #feed-post-N anchor for traceability.
  • new (people): alon-huri — first thought-leader entry; weekly Friday LinkedIn poster.
  • new (companies): team8 — second VC entry after entree-capital; both tagged investor-thesis.
  • new (sources, 5 wiki source pages out of 15 raw): 2026-05-09-alonhuri-linkedin-ai-native-company, 2026-05-16-alonhuri-linkedin-ai-native-growth-hacking, 2026-03-06-alonhuri-linkedin-saas-is-dead-ai-native-agency, 2026-01-30-alonhuri-linkedin-full-stack-ai, 2026-01-09-alonhuri-linkedin-vibe-acting.
  • new (topics): full-stack-ai-vertical-services — synthesis topic anchored by the AI-Native-Agency + Full-Stack-AI source pair; reads as the most aggressive cut of vertical-use-case-led-brain.
  • updated: context-os-brain (source_count 2→3; new Evidence row from ai-native-company post), agent-native-go-to-market (source_count 3→5; +2 Evidence rows from growth-hacking + vibe-acting posts), vertical-use-case-led-brain (source_count 3→5; +2 Evidence rows from saas-is-dead + full-stack-ai posts; Related extended with full-stack-ai-vertical-services), wiki/index.md (1 People row, 1 Companies row, 5 Sources rows, 1 Topics row), wiki/log.md (this entry).
  • Curation gate (load-bearing for ideation? validate / contradict / extend a wiki topic OR open a new vertical/wedge angle?): 5 of 15 posts passed → wiki source page. 10 of 15 stayed raw-only with the rationale documented inline on alon-huri’s ## Notes — non-wiki posts (raw only) section. Load-bearing breakdown: ai-native-company (verbatim with context-os-brain), ai-native-growth-hacking (extends agent-native-go-to-market), saas-is-dead-ai-native-agency + full-stack-ai (anchor new topic full-stack-ai-vertical-services as most-aggressive cut of vertical-use-case-led-brain), vibe-acting (sharpens agent-native-go-to-market with action-layer thesis). Skip rationale: tactical playbooks (creator-arbitrage, reddit-personal-brand, reddit-experiment-results), personal/social initiative (soldiers-ptsd × 2 — important to Alon, not to ideation), opinion / speculation (ads-in-ai-chats, vibe-coding-2 / ASI, apple-openai-merger), founder-finance philosophy (bootstrap-vs-astrap — relevant only when team raises), methodology overlap with team’s own (100-ideation-hacks).
  • Skepticism flag. alon-huri is Managing Partner at team8. Five of his five wiki-promoted posts agree very strongly with the team’s pre-existing strategic direction. That’s more suspicious, not less — Alon is plausibly pitching narratives his fund’s portfolio plays into. Each source page carries this caveat inline; future-Claude should not treat his agreement as confirmatory without cross-checking against non-VC voices.
  • notes: No action items generated — these are external thought-leadership posts, not meetings; no follow-ups owned by team members. Open question worth raising at the next directions sync (flagged on full-stack-ai-vertical-services): of Alon’s 15 named regulated-profession verticals (legal, real-estate, finance, healthcare, education, marketing — see 2026-03-06-alonhuri-linkedin-saas-is-dead-ai-native-agency), only healthcare (ari-leshno / clinical-data-portability) has been concretely evaluated by the team; the other 12 remain unfiltered. Filename convention. Raw files use YYYY-MM-DD-alonhuri-linkedin-<topic-slug>.md per user request — date is post-creation date (approximated to nearest Friday for the activity-feed sweep, since LinkedIn shows only relative time like “3mo”). Source-page slugs match the raw filenames for 1:1 navigation. Memory updated. Project memory at .claude/projects/.../memory/brain-purpose-startup-ideation.md already covered the VC-perspective input pattern from the entree-capital ingest; no additional memory needed for this Alon ingest.

[2026-05-31] ingest | meeting: 2026-05-31-directions-connect-businesses-and-brain-scaling

  • source: raw/meetings/notetaker/2026-05-31-directions-starting-the-week-17-00.json (Hebrew ASR transcript, ~90 min; opening ~5 min football championship + closing ~10 min logistics + assorted personal asides excluded from summary)
  • new: 2026-05-31-directions-connect-businesses-and-brain-scaling, connect-businesses-multi-vertical-api (NEW topic), plaid, snaptrade, papaya-global, wolt, rami-levy, victory, osher-ad, render-ai (8 new companies)
  • updated: saar-arbel (source_count 3→4), guy-barkat (source_count 3→4), nizan-shifman (source_count 2→3), tomer-shalom (source_count 1→2), alon-huri (source_count 5→6; new mentioned-in row for the Saar Claude-project ideation use), etoro (source_count 2→3; partnership pipeline + NoteTaker policy), wonderful (source_count 2→3; ARR + FDE-debate update), sierra (source_count 2→3; ~$1B raise per Nizan), context-os-brain (source_count 3→4; agent-harness + difficulty-as-moat + qualified-source-of-truth additions; updated open questions), vertical-use-case-led-brain (source_count 5→6; vertical-determines-data-model evidence; sibling non-Brain directions called out), prediction-markets-regulatory-play (source_count 1→2; triage-gate escalation evidence row), wiki/index.md (8 new Companies, 1 new Topic, 1 new Meeting), wiki/action-items.md (1 prior item closed — Sunday-evening sync was today; 1 prior item moved Open → In progress — Toto chairman follow-up; 6 new open items + 1 progress note across 4 prior items)
  • notes: NEW direction surfaced. guy-barkat introduced a third active exploratory direction — connect-businesses-multi-vertical-api — alongside vertical-use-case-led-brain and prediction-markets-regulatory-play. Crucially this is a non-Brain direction; the team now holds three open lenses without a process for converging. Brain reframe. Same meeting deepened the Brain conversation: Guy’s strongest scaling-difficulty case yet (partial data is worse than no data; conflict resolution unsolved; need an “agent harness”) flipped mid-conversation into a moat argument (the difficulty itself prevents incumbents from building it in-house). Saar reinforced vertical-as-data-model. Nizan defended Holy-Grail framing and reported “I’ve cracked AI” via daily Brain use. Source-of-truth debate resolved as “qualified yes when paired with conflict-surfacing UX.” Adoption gap. guy-barkat still unattached to the Brain; provisioning + pair-install scheduled for this week. etoro’s default opt-out on Monday NoteTaker is now documented as the policy gate. No decision pages. Two-vertical-direction ramping + Brain harness work are still exploratory; no commitments crossed the decision-slug threshold. Note on Yoni. The new meeting references “our CEO Yoni” at etoro (Yoni Assia) — different person from the existing yoni page (Monday senior exec / pitch target). No new Yoni page created since the eToro Yoni was only mentioned in passing; a future ingest with substantive Yoni-Assia content should disambiguate via Section C collision-suffix (yoni-assia vs yoni-monday rename).

[2026-06-04] ingest | meeting: 2026-06-04-directions-matan-construction

  • source: raw/meetings/notetaker/2026-06-04-directions-w-matan-construction-06-00.json (Hebrew ASR transcript, ~31 min; expert interview format — Matan was on the road, so the call was kept tight)
  • new: 2026-06-04-directions-matan-construction, matan-amram, solel-boneh, akrstein, construction-project-management
  • updated: nizan-shifman (source_count 4→5; new mentioned-in entry), guy-barkat (source_count 5→6; new mentioned-in entry), wiki/index.md (1 Person, 2 Companies, 1 Meeting, 1 Topic added), wiki/action-items.md (1 item closed to Done — construction intro was Matan himself; 3 new Open items for follow-up contacts)
  • notes: First construction-vertical expert interview. Key finding: internal project management at even the largest Israeli GCs is entirely offline Excel + WhatsApp — no live visibility. Two emerging angles: (1) “context OS for construction” — Nizan’s thesis that project-state visibility is the root problem; (2) the subcontractor/labor-deployment layer — Guy’s thesis that this is an equally underserved, entirely un-tooled segment. No decision pages created — still mapping the space. Entity note: akrstein has multiple ASR spellings in the transcript (אקרשטיין / הקירשטיין / האקלשטיין); canonical slug is akrstein with aliases covering variants.

[2026-06-02] ingest | meeting: 2026-06-02-directions-account-management-vertical

  • source: raw/meetings/notetaker/2026-06-02-directions-brainstorm-17-30.json (Hebrew ASR transcript, ~55 min; opening ~3 min personal chitchat dropped)
  • new: 2026-06-02-directions-account-management-vertical, daniela, ofer-polivoda, nir-goldstein, mixpanel, gong, account-management-vertical
  • updated: nizan-shifman (source_count 3→4), guy-barkat (source_count 4→5), saar-arbel (source_count 4→5), harmony (source_count 1→2), vertical-use-case-led-brain (source_count 6→7; “which vertical” open question resolved to AM working answer), context-os-brain (source_count 4→5; AM architecture note + agentic CRM debate), wiki/index.md (3 People, 2 Companies, 1 Meeting, 1 Topic added), wiki/action-items.md (2 items closed to Done; 4 new Open items)
  • notes: The team’s “which vertical” open question has a working answer for the first time: Account Management / Customer Success. No formal decision page created — still pre-commitment, with the Daniela pilot as the next validation gate. Key new entities: daniela (Harmony AM, pilot subject), nir-goldstein (VC framing: grow book-of-business coverage), ofer-polivoda (AM benchmark + feedback target; originally captured as Ofer/Mixpanel before full name was known), gong (revenue-intelligence adjacent). Architecture debate (integrate CRMs via MCP vs. build AM-native CRM) is live and open — no decision taken.

[2026-05-31] lint | post-ingest checklist

  • counts: 71 wiki files (3 operational root + 68 slugged content). Index references 69 distinct slugs. Frontmatter present + complete on all 68 content pages. No frontmatter wikilinks. No source_count:0 orphans. No status:active pages older than 90 days. No Open/Done duplicates in action-items.md. Action-items.md has all four required H2 sections in order; every Source meeting reference resolves.
  • new flagged + auto-fixed in this run:
  • pre-existing flagged (NOT auto-fixed; out of scope for this lint run):
    • [[pagaya]] referenced from companies/axonius.md, meetings/2026-05-27-directions-vertical-pivot-and-prediction-markets.md, people/guy-barkat.md — no page exists. Pagaya is a real Israeli AI-fintech; intro pipeline reference. Suggested fix: create companies/pagaya.md stub on next refactor.
    • [[maccabi-netanya]] referenced from people/tomer-shalom.md — no page exists. Football club; lower priority. Suggested fix: create stub or unlink.
    • [[decision-slug]] flagged from meetings/2026-05-06-brain-os-strategy-brainstorm.md and meetings/2026-05-27-directions-vertical-pivot-and-prediction-markets.md — false positives; both occurrences are inside <!-- ... --> HTML template comments (template scaffolding from Section E meeting template), not body wikilinks. Lint script doesn’t strip HTML comments. Suggested fix: strip the template-comment lines on next refactor of those meeting pages, OR upgrade the lint regex to skip HTML comments.
    • [[slug]] flagged from index.md — same false-positive pattern; the Decisions section header is <!-- [[slug]] — one-line summary (date + outcome) -->. Same fix options.
  • result: PASS for this ingest (all 3 ingest-introduced issues fixed); 4 pre-existing flags surfaced for future cleanup.

[2026-06-08] ingest | source: 2026-06-08-contech-funnel-brief

[2026-06-11] query | Almogim construction-Brain POC design

  • source: user conversation 2026-06-11 (Saar ↔ Almogim CEO; no raw transcript yet — flagged on the new pages)
  • new: almogim, almogim-exec-visibility-poc
  • updated: construction-margin-protection-wedge (second access path), construction-project-management (related link), wiki/index.md
  • notes: exec-visibility POC scoped (M365/Outlook, concierge model, Hebrew chat) with week-4 pivot gate to the margin wedge based on CEO query-log classification. source_count: 0 on both new pages until a raw source (meeting summary or first working session) is captured.

[2026-06-14] ingest | meeting: 2026-06-13-directions-revenue-cycle-and-construction-poc

[2026-06-14] refactor | OKF-compatible wiki frontmatter

  • reason: User requested support for Google’s Open Knowledge Format frontmatter structure.
  • touched: all non-operational wiki/**/*.md concept pages; AGENTS.md; CLAUDE.md; log
  • notes: Added OKF recommended title, description, and timestamp fields to 98 wiki concept pages. Added resource for meeting/source pages where a canonical raw path or source URL exists. Preserved existing repo fields (aliases, source_count, last_updated, status, etc.) as OKF producer-defined extension keys. Left raw/ immutable and left operational root files (wiki/index.md, wiki/log.md, wiki/action-items.md) without frontmatter.

[2026-06-15] ingest | meeting: 2026-06-15-directions-ofer-polivoda-am-operator-interview

[2026-06-15] refactor | define local OKF profile

  • reason: User approved the OKF-compatibility plan: keep wiki/ as the OKF bundle root, preserve repo-native wikilinks internally, and verify every concept page follows the local frontmatter profile.
  • touched: AGENTS.md, CLAUDE.md, README.md, scripts/okf-profile-lint.rb, log
  • notes: Added explicit OKF bundle boundary, documented wiki/action-items.md as a local operational extension, kept strict Markdown-link conversion as an export-time concern, and added ruby scripts/okf-profile-lint.rb. Verification passed for 99 concept pages.

[2026-06-15] refactor | replace OKF profile lint with Python

  • reason: User clarified that repo code should be Python-only.
  • touched: scripts/okf_profile_lint.py, scripts/okf-profile-lint.rb, AGENTS.md, CLAUDE.md, README.md, .gitignore, log
  • notes: Replaced the Ruby OKF-profile lint script with a Python/PyYAML implementation, ignored Python bytecode caches, and updated current docs to use python3 scripts/okf_profile_lint.py. Verification passed for 99 concept pages.

[2026-06-16] ingest | meeting: 2026-06-16-directions-ran-levy-fiverr-cs-operator-interview

[2026-06-16] ingest | meeting: 2026-06-14-directions-from-home-strategy-brain-dogfooding

[2026-06-16] ingest | meeting: 2026-06-16-directions-ofri-avivi-renewals-operator-interview

[2026-06-16] query | renewals-management-wedge synthesis

[2026-06-20] ingest | notetaker enrichment: 2026-06-16-directions-ofri-avivi-renewals-operator-interview

  • source: raw/meetings/notetaker/2026-06-16-directions-w-ofri-avivi-09-00.json
  • updated: 2026-06-16-directions-ofri-avivi-renewals-operator-interview, renewals-management-wedge, account-management-vertical, ofri-avivi, cymulate
  • notes: Enriched the existing Ofri/Cymulate renewals pages from monday Notetaker summary + topics; no new independent source_count because this is the same meeting as the legacy raw transcript. Added sharper details on GRR vs NRR scope, Salesforce+Coda manual renewal workflow, BI removal during CRO/tool-stack change, CS joining late-stage sales calls, partner-mediated renewal signal loss, and technical fluency as a renewal-manager scaling constraint. No action items were captured by Notetaker.

[2026-06-20] lint | post-notetaker enrichment check

  • counts: 107 concept pages
  • result: python3 scripts/okf_profile_lint.py passed
  • flagged: none

[2026-06-21] ingest | meeting: 2026-06-20-directions-planning-the-week-18-30

[2026-06-21] lint | post-planning-meeting ingest check

  • counts: 114 concept pages
  • result: python3 scripts/okf_profile_lint.py passed
  • flagged: none

[2026-06-21] ingest | source: 2026-03-23-a16z-two-paths-left-for-software

[2026-06-21] lint | post-a16z-source ingest check

  • counts: 118 concept pages
  • result: python3 scripts/okf_profile_lint.py passed
  • flagged: no new ingest-introduced issues. A separate ad-hoc wikilink scan still surfaces pre-existing unresolved links already documented in earlier lint notes: pagaya, maccabi-netanya, and template placeholders such as decision-slug.

[2026-06-21] refactor | introduce VC hypothesis layer

[2026-06-21] lint | post-hypothesis-layer refactor check

  • counts: 123 concept pages
  • result: python3 scripts/okf_profile_lint.py passed
  • flagged: none from the local OKF profile check.

[2026-06-21] ingest | source: 2026-03-05-sequoia-services-the-new-software

[2026-06-21] lint | post-Sequoia-source ingest check

  • counts: 130 concept pages
  • result: python3 scripts/okf_profile_lint.py passed
  • flagged: none from the local OKF profile check.

[2026-06-26] refactor | organize Notetaker state and gate digests

  • reason: User asked to clarify the folder structure and digest only monday Notetaker meetings where at least two of Guy, Nizan, and Saar participated.
  • touched: AGENTS.md, CLAUDE.md, README.md, skills/fetch-notetaker-meetings/SKILL.md, skills/update-directions-brain-context/SKILL.md, skills/fetch-notetaker-meetings/scripts/fetch_notetaker_meetings.py, state/notetaker/processed.md
  • notes: Moved mutable Notetaker processing state from top-level notetaker/processed.md to state/notetaker/processed.md; documented raw/ as immutable inputs, wiki/ as synthesized knowledge, and state/ as mutable pipeline bookkeeping. The recurring Notetaker sync now fetches participant metadata, saves only meetings with at least two of guybarkat2@gmail.com, nizanshifman@gmail.com, and saararbel1@gmail.com, and records others as skipped-not-core-team so they are not repeatedly digested.

[2026-06-26] refactor | organize raw sources

[2026-06-26] ingest | meeting: 2026-06-21-directions-roy-keren-salesforce-gtm-ai-interview

[2026-06-26] lint | wiki profile check

  • counts: 133 concept pages
  • result: python3 scripts/okf_profile_lint.py passed
  • flagged: none

[2026-07-16] query | Alon Huri top 3 LinkedIn theses

[2026-08-01] query | Insurance 2.0 — SMB insurance competitive landscape (Alon Huri direction)

  • source: user-pasted Gemini chat export (NOT saved to raw/ per user choice — context only) + 3 parallel web-research passes (Aug 2026)
  • new: smb-insurance-portfolio-brokerage, smb-insurance-competitive-landscape, ai-native-brokerages-capture-smb-distribution, next-insurance, coverwatch, harper, embroker, coverdash, clark
  • updated: alon-huri, team8, full-stack-ai-vertical-services, wiki/index.md
  • notes: Direction originates from a ~July 2026 VC ideation session with Alon Huri (user-reported; no transcript in raw/ — claims from it marked as such on pages). Key corrections to the Gemini-sourced picture: (1) NEXT Insurance WAS acquired — Munich Re/ERGO, 4.5M pre-seed, July 2026) occupies the exact wedge; (3) a 2025–26 AI-native brokerage wave exists that Gemini missed entirely (Harper 175B small-commercial TAM (Munich Re), 87.7% IA channel share (Big “I” 2025), 77% SMBs underinsured (Hiscox 2025), JD Power renewal intent at multi-year low. Team8 check: Sayata is their only insurtech bet; no announced foundry insurance build, but Alon is publicly ideating in insurance (CTech Apr 2025; Team8 “Designing 2026” Vibe-Shopping thesis). Company-stub scope stayed selective: 6 load-bearing pages; the long tail (Pie, Coterie, Corgi, Kinro, Panta, Qumis, etc.) lives inside the landscape page until load-bearing.

[2026-08-04] ingest | meeting: 2026-08-03-directions-smb-insurance-value-chain-mapping

[2026-08-04] ingest | meeting: 2026-08-03-directions-tom-granit-lemonade-car-insurance-interview

[2026-08-04] ingest | meeting: 2026-08-03-directions-shai-slobodov-lemonade-carrier-interview

[2026-08-04] ingest | meeting: 2026-08-03-directions-round-2-super-broker-vision

[2026-08-04] lint | post-insurance-offsite ingest check

  • counts: 153 concept pages
  • result: python3 scripts/okf_profile_lint.py passed
  • flagged: wikilink scan over all touched pages found no new unresolved links; pre-existing known unresolved link pagaya (in guy-barkat) remains, documented in earlier lint notes.

[2026-08-07] query | F2F learnings: AI-brokerage scope locked + Corgi category-creation GTM

  • source: user-reported F2F team session (WhatsApp summary shared in Cowork session; no transcript in raw/)
  • new: (none)
  • updated: smb-insurance-portfolio-brokerage
  • notes: Scope decisions recorded: AI brokerage for end customers (internal + customer-facing process optimization); competitor set = coverwatch + harper; roll-up path rejected (heavy ops, mismatched books); “AI for brokerages” tooling rejected (small derivative TAM). New GTM thesis from the Corgi pattern: pick segment via industry × use-case cross, package existing policies into a unique named coverage, position as the category-defining broker for a new risk. Open questions: beachhead segment, exact processes, GTM, differentiation. Learnings also mirrored to the external research-hub artifact (outputs/insurance-research-hub.html).

[2026-08-07] ingest | meeting: 2026-08-07-directions-nizan-saar-ai-brokerage-thesis-and-pitch

[2026-08-07] lint | post-2026-08-07-session ingest check

  • counts: 154 concept pages
  • result: python3 scripts/okf_profile_lint.py passed
  • flagged: no new unresolved wikilinks; pre-existing known pagaya link (in guy-barkat) remains.

[2026-08-07] query | promote 2026-08-07 session assumptions to hypothesis layer

[2026-08-07] ingest | Cowork research sessions: distribution economics, TAM model, segment sprint, Corgi decomposition

  • source: Cowork research sessions Aug 1–7 (web research passes; artifacts in session outputs: insurance-research-hub.html, competitive quadrant, learning deck, strategy memo). No raw transcript — analytical layer, complements the 2026-08-07 Notetaker-based meeting ingest.
  • new: us-insurance-distribution-economics, ai-brokerage-tam-model, insurance-beachhead-segment-sprint
  • updated: corgi (success decomposition, not-first lineage, momentum caveat, avoid-their-network implication), smb-insurance-portfolio-brokerage (links to the analytical layer), wiki/index.md
  • notes: Key filed conclusions: (1) money flows through carriers but sticks at distribution — broker P&L is 57% attackable human labor (18–26B recurring commission pool but only ~420M revenue ≈ NEXT’s exit revenue, implying ~330 audits/day (a product requirement, not just a growth number); (3) segment sprint scored 8 crosses — robotics/automation integrators won 22/25 (CG 35 08 completed-ops AI exclusion + contract-requirements shock + A3/CSIA reachability + unowned name), with 4 diligence kill-gates before commitment; per-segment pools are beachheads, not markets. ⚠-marked derived figures throughout; data holes named (account-band premium split, audit→BoR conversion, US switching rates).

[2026-08-07] refactor | add wiki/axioms.md — idea-evaluation constraint registry

  • reason: Team decision (Saar): maintain a living registry of decided constraints (“axioms”) + evidence-derived heuristics so every future AI session applies accumulated critical thinking when evaluating ideas.
  • touched: wiki/axioms.md (new, fourth operational root file), CLAUDE.md (invariants 4–6, Section B root-files list + tree, Section H query workflow steps 2 and 6, Section I lint checklist), scripts/okf_profile_lint.py (OPERATIONAL_FILES set)
  • notes: Seeded with the insurance direction: 6 axioms (AX-INS-1..6: not-a-carrier; AI brokerage for end customers; no roll-ups; no broker/carrier tooling; category-creation differentiation; premium-is-never-TAM) each with rationale, source wikilink, and revisit trigger; 9 heuristics (H-INS-1..9: thin accounts kill brokerages, parsing isn’t a moat, BoR = renewal-window campaign, beachhead ≠ market, avoid Corgi’s home field, embedded = experiment, five-part segment filter, unproven-until-renewal-cycle, soft-market discipline) each citing evidence pages. Governance: entries change only via logged team decisions; retired entries preserved. Invariant 6 added: read axioms.md before any idea evaluation and classify ideas as consistent/violates/challenges with explicit axiom IDs.

[2026-08-08] query | backfill post-session WhatsApp learnings (2026-08-07 evening)

  • source: WhatsApp thread Saar↔Nizan, 2026-08-07 17:10–18:09 (screenshots shared in Cowork session; user-reported, no raw capture)
  • updated: insurance-beachhead-segment-sprint, alon-huri, wiki/index.md
  • notes: Coverage check found the thread mostly restates the already-ingested 12:00 session. Two gaps backfilled at user request: (1) a second, unscored segment-search axis — existing industries where all incumbents are old-school (example: restaurants insurance) — added to the sprint page as a key point + open question on whether it deserves its own scored pass or is dominated by the category-creation play; (2) Huri pitch-calibration expectation — he wants to see the team understands the meaning of full-stack and of focusing on a specific coverage + segment — added to the alon-huri page with the deck implication (explicitly answer why broker-not-carrier now and when that flips). Deliberately NOT filed (user deferred): Faye as a full-stack comparable stub; the fragmentation→consolidation reframe of the Corgi story.

[2026-08-09] lint | weekly Notetaker sync + OKF profile lint

  • source: /update-directions-brain-context weekly sync (—since 7d)
  • sync: 5 meetings matched; 4 already processed; 1 new meeting “Meeting with Timor” (2026-08-09) marked skipped-not-core-team in state/notetaker/processed.md; 0 new full fetches, nothing to ingest
  • lint: scripts/okf_profile_lint.py passed, 162 concept pages
  • flagged: none

[2026-08-09] ingest | meeting: 2026-08-09-directions-timor-arbel-sadras-growth-vc-advisory

[2026-08-09] query | is there a US equivalent of Cover (Israel)?

  • trigger: Saar asked whether a US product like my.coverai.co.il exists. Short answer: no single equivalent — the value prop is split across Marble (acquired by The Zebra, July 2024), Jerry and Coverage Cat (personal lines), Capitalize/Beagle (401k side), and Canopy Connect/Axle (data-access infrastructure).
  • new: cover, policy-ingestion-requires-a-data-rail
  • updated: clark, smb-insurance-competitive-landscape, wiki/index.md
  • web-verified 2026-08-09: cover.co.il homepage (Pension Clearing House + Ministry of Finance data access; 350K+ users; free); Adif+ (Cover recruiting dozens of licensed pension agents; co-founder/COO Lior Shekory); Capital Market Authority Har HaBituach portal harb.cma.gov.il + harhabituach.co.il (operator = Capital Market, Insurance and Savings Authority; scope = motor/home/health/LTC/mortgage-life; insurers required to report; commercial insurance excluded); usecanopy.com + einpresswire (Canopy Connect $6.5M Series A, Nevcaut Ventures); axle.insure. No raw capture taken — analytical layer, consistent with the 2026-08-01 competitive-research precedent.
  • notes: The interesting finding is structural, not competitive. Cover and clark can give a portfolio audit away free because a regulator (Israel) or a standing mandate (Germany, unverified) supplies the data; the US has no registry at any level, so ingestion degrades into customer-supplied documents plus commoditized parsing. Filed as policy-ingestion-requires-a-data-rail at medium confidence — the mechanism is well-evidenced on the railed side, but every railed data point is personal lines/pension while the team’s direction is SMB commercial. It reinforces AX-INS-5 and H-INS-2 with a structural argument, and predicts the document handoff (not the analysis) is the real US funnel gate. Correction logged: the earlier conversational answer implied Marble was still an independent going concern; smb-insurance-competitive-landscape already recorded the Zebra acquisition.
  • proposed, NOT added (requires a team decision per axioms governance): a heuristic H-INS-10 — “In the US there is no policy-data rail; portfolio ingestion is a funnel gate, not free acquisition” for wiki/axioms.md.

[2026-08-10] ingest | meeting: 2026-08-09-directions-operational-ai-and-axiom-challenge

  • source: raw/meetings/notetaker/2026-08-09-directions-inssurance-14-30.json
  • new: 2026-08-09-directions-operational-ai-and-axiom-challenge, goldfus, canopy-connect
  • updated: smb-insurance-portfolio-brokerage, smb-insurance-competitive-landscape, insurance-beachhead-segment-sprint, ai-collapses-smb-brokerage-labor-cost, policy-ingestion-requires-a-data-rail, saar-arbel, nizan-shifman, guy-barkat, index, action-items
  • notes: Evening session of 2026-08-09. Central reversal: the team concluded most AI disruption today sits in internal broker/MGA operations rather than end-customer personalization, reaffirming “brokerage is fundamentally a labor business” (60-70% operations; >50% of cost is labor) and identifying the AMS (not CRM) as the system of record. Best product concept: “Gong for insurance” — broker call recording to structured AMS data plus proactive coverage actions. NOTABLE: that concept sits inside the category AX-INS-4 forbids, and the axiom was openly contested in-session (“one of the axioms I’m not sure I agreed to”); AX-INS-2 is strained by the same reversal plus a new pension/heavy-lines axis. No decision was taken, so per the axioms protocol NO axiom was amended — the tension is recorded on the meeting page, the direction page, and as an action item for an explicit decision session. Also: YC scan of ~15 insurance startups (“everyone wants to be a brokerage”) surfaced RiskyTix (?) spanning robotics/autonomy/data-centers/energy, partly pre-empting the segment-sprint winner; goldfus surfaced as first design-partner candidate; canopy-connect independently identified as the closest US analog to Israel’s mandated rails, converging with policy-ingestion-requires-a-data-rail. Company names from the YC scan are Hebrew-transcript transliterations, marked (?) pending verification. Core-team gate override: only Saar’s account joined; user invoked the skill to process unprocessed meetings.

[2026-08-10] lint | post-2026-08-09-session ingest check

  • counts: 174 concept pages
  • result: python3 scripts/okf_profile_lint.py passed
  • flagged: index coverage complete (every concept page appears exactly once). No new unresolved wikilinks; pre-existing known ones remain: pagaya, maccabi-netanya, and template placeholders (decision-slug, slug) inside historical log entries.

[2026-08-10] query | US licensing mechanics + first Europe/LatAm geography test

  • trigger: Saar asked (a) how long a producer licence takes per US state and what the process is, and (b) whether the team ever considered running the Harper/CoverWatch play in Europe or LatAm.
  • new: us-producer-licensing-mechanics, geography-alternatives-europe-latam, carrier-access-not-licensing-gates-new-brokerages
  • updated: clark (rail mechanism resolved; €1B → €613M markdown; status stub → active), policy-ingestion-requires-a-data-rail (named Clark disconfirming test RUN and resolved toward the narrow reading; FIDA added), wiki/index.md
  • headline finding: licensing is not a barrier and carrier appointments are. 49 states + DC reciprocal, ~30 states in 3 weeks, 48–50 in 10–14 weeks, ~100–250K committed premium each. Filed as a hypothesis because it implies the barrier and the moat are the same object.
  • CONTRADICTION SURFACED: 2026-08-03-directions-shai-slobodov-lemonade-carrier-interview records shai-slobodov saying Europe’s unified EU regulation is “a legitimately easier arena.” Research contradicts this — US 50-state licensing is faster and cheaper than a single EU authorisation (105-day average in Ireland; €1.56M mandatory PI; German-language exam), and EU passporting is used by only 7,269 of 685,846 registered intermediaries (~1%), declining, with EIOPA unable to measure whether any business flows. The meeting page was deliberately NOT edited — it is an accurate record of what he said; the correction lives on geography-alternatives-europe-latam.
  • LatAm ruled out on substrate, not difficulty: ~17–18% of Mexican and ~26.7% of Brazilian SMBs carry any business insurance, so the thesis inverts from optimisation/switching to first-sale/penetration. Brazilian commission attaches to the signing broker (Lei 4.594/64 art. 13) so BoR does not monetise mid-term; Mexican cover lapses automatically 30 days after non-payment (LCS art. 40). Exit comparables fatal (Wiz Co: R1.36B market cap).
  • UK flagged as the only genuine alternative geography (94% broker share of commercial; AR route live in 2–3 months) and the ABI/Public First Jan-2026 SME survey as the best quantified evidence anywhere for the continuous-optimisation gap (50% had not reviewed cover in 12+ months while 52% had a material business change).
  • NOTE: no axiom amended. This research materially strengthens the case that AX-INS-5’s “category creation” is an entry tactic rather than a moat — already flagged for an explicit decision session, still pending a team decision.

[2026-08-10] lint | post-geography/licensing-filing check

  • counts: 177 concept pages (174 → 177)
  • result: python3 scripts/okf_profile_lint.py passed
  • flagged: index coverage complete (every concept page appears exactly once); all wikilinks in the three new pages and the two edited pages resolve; required frontmatter present on all new pages. Pre-existing known unresolved links remain unchanged (pagaya, maccabi-netanya, historical template placeholders).

[2026-08-15] ingest | meeting: 2026-08-10-directions-smb-survey-verticals-and-fida

  • source: raw/meetings/notetaker/2026-08-10-test-10-51.json
  • new: 2026-08-10-directions-smb-survey-verticals-and-fida, wynter
  • updated: geography-alternatives-europe-latam, policy-ingestion-requires-a-data-rail, insurance-beachhead-segment-sprint, alon-huri, saar-arbel, guy-barkat, nizan-shifman, index, action-items
  • notes: Notetaker title “test” (10:51) is a real morning working session — all three founders speaking; participant metadata empty, gate overridden per explicit user request to process this meeting ID. Content: SMB segmentation by size + bundling difficulty; three survey verticals (IT services 10–50 employees fully digital, accounting/tax D2C-closable, restaurants online-provable); Wynter cost anchor (1,250 credits/250 respondents, new survey 30–50); the team consumed the same-day geography/licensing research in-session (LatAm struck at 17% insured — Notetaker summary misattributed the figure to the UK, transcript-verified as לט”ם; UK kept as only real second market); Guy’s contrarian FIDA-2029 pre-positioning idea filed on the data-rail hypothesis; BOP fluency drill for the Huri meeting.

[2026-08-15] ingest | meeting: 2026-08-10-directions-huri-pitch-deck-build

[2026-08-15] ingest | meeting: 2026-08-13-directions-fundraising-and-validation-plan

  • source: raw/meetings/notetaker/2026-08-13-directions-08-00.json
  • new: 2026-08-13-directions-fundraising-and-validation-plan, insurance-seed-fundraising, maor-friedman, f2, boaz
  • updated: alon-huri, team8, brm, goldfus, avner, cover, saar-arbel, guy-barkat, nizan-shifman, index, action-items
  • notes: Guy+Saar session (Nizan in participant metadata, not an active speaker). Central reframe via Boaz: the team is effectively mid-round with Huri/Team8 (fund CRM logs meetings; informal foundry-style ideation); next Huri conversation likely decisive; Huri-personal-only investment flagged as negative signal. Two-track validation before that meeting: broker blitz Monday + free design-partner “transformation” (US preferred, Avner/Goldfus fallback) and ~₪5K Cover-style B2C demand experiments. Cover-style B2C named “the biggest prize” — tension with AX-INS-2 (B2B SMBs) surfaced on the meeting page, no axiom decision taken. New fundraising topic page opened to track round state.

[2026-08-15] ingest | meeting: 2026-08-15-directions-acquisition-wedge-and-deck-structure

[2026-08-15] lint | post-ingest check for the four 2026-08-10..15 meetings

  • counts: 189 concept pages (177 → 189; 12 new: 4 meetings, 1 decision, 1 topic, 3 people, 3 companies)
  • result: python3 scripts/okf_profile_lint.py passed
  • flagged: index coverage complete; all wikilinks in new/edited pages resolve. Pre-existing known unresolved links unchanged (pagaya, maccabi-netanya, historical template placeholders in old log entries/meeting pages).
  • processed-registry: all four raw JSONs marked processed in state/notetaker/processed.md.

[2026-08-15] refactor | axiom registry updated to match the 2026-08-10..15 sessions (user-directed team decision)

  • reason: Saar instructed “change the axioms according to what we said lately” — treated as the explicit team decision required by axioms governance.
  • touched: wiki/axioms.md, wiki/action-items.md (closed the AX-INS-2/AX-INS-4 decision-session item)
  • changes:
    • AX-INS-2 amended: B2C = later expansion prize (2026-08-13); phase-1 execution via a partner brokerage while we own brand/traffic/customer; new revisit trigger for partner book-capture (2026-08-15-acquisition-wedge-before-workflows).
    • AX-INS-4 reaffirmed: the 2026-08-09 challenge resolved by the 2026-08-15 rejection of the FDE-for-brokers path.
    • AX-INS-5 RETIRED (moved to Retired, never deleted): category-creation differentiation superseded; honesty note that its revisit trigger was never formally run.
    • AX-INS-7 added: differentiation = aligned incentives (fixed fee + neutral comparative rater) + genuine post-bind continuity at the generalist+continuous corner.
    • AX-INS-8 added: prove acquisition before workflows; rent execution via partner brokerage; no FDE-for-brokers.
    • AX-INS-9 added: US-first, UK the only second market, LatAm off the table; FIDA is 2029+ option value.
    • H-INS-10 adopted (proposed 2026-08-09): no US policy-data rail — portfolio ingestion is a funnel gate with its own conversion kill-gate.
  • notes: sources for every entry cite the decision/meeting pages ingested earlier today. Hypothesis pages (named-coverage-category-creation-wins-gtm) intentionally NOT retired — the hypothesis remains falsifiable and tracked; only its axiom status changed.

[2026-08-15] query | tech & integrations blueprint for the AI brokerage (Saar’s deck workstream)

[2026-08-15] query | the “aggregator” layer: Bold Penguin and the market-access competitors

  • source: web research (Aug 2026), two parallel passes — Bold Penguin teardown and the four-category aggregator map; load-bearing claims verified against boldpenguin.com/about, the Exchange product page, and acquisition press (iBynd/Semsee Mar 2026; Penn-America/Sayata Sept 2025).
  • new: insurance-distribution-aggregator-layer, bold-penguin
  • updated: insurance-tech-integration-stack, index, outputs/tech-blueprint-integrations-2026-08.md (two corrections)
  • notes: Answers Saar’s question on market aggregators. Core disambiguation: “aggregator” = four businesses (agency networks à la SIAA/Smart Choice — what the industry actually means; comparative raters; digital wholesalers; consumer comparison sites). Bold Penguin is a rater+lead-marketplace that bought a wholesaler (SquareRisk, Oct 2025) and now states its vision is “the premier digital broker for commercial insurance” — the tooling→brokerage ladder run by an American Family-owned company; founder Bodner replaced Jan 2025 by AmFam’s own CSTO, and the 2021 neutrality pledge no longer appears on the site. Its Exchange runs on human CSM phone-verified warm transfers — the labour structure an AI-native funnel undercuts. Key new pattern: every Category-B rater reaching distribution relevance was acquired by a carrier/MGA/AMS 2021–26 (Bold Penguin→AmFam, Tarmika→Applied, Talage→Mission, Sayata→Penn-America, Semsee→iBynd); only Herald and CoverForce remain independent — validates AX-INS-7’s neutrality wedge as structurally scarce, and flags our rented rails as acquisition targets. Corrections applied to the tech blueprint (Semsee no longer independent; Sayata acquired, not merely in sale process). New diligence item raised for Guy’s partner-brokerage search: ask whether the candidate belongs to an agency network, which may hold a commission override or vested interest in the book — an unexamined threat to the AX-INS-2 ownership question. Contrary evidence filed against the AI-disintermediation narrative (188 insurtech applications for 2026: mostly tools for intermediaries, “minimal disintermediation focus”) — consistent with H-INS-8. Margin-migration tell: Ryan Specialty FY25 Underwriting Management +58.5% vs Wholesale Brokerage +7.5%.

[2026-08-16] query | check-up funnel implementation design (spec + clickable prototype)

  • source: Saar’s question “how can I implement this product flow” (signup → carrier auth → download → analyse). Research: Canopy Connect developer docs (docs.usecanopy.com, verified endpoint/webhook/schema level) + US commercial document-type analysis (ACORD 25 2016/03 specimens, a live Chubb BOP dec page, ISO form numbering).
  • new: insurance-checkup-funnel-design
  • updated: index; deliverables outputs/checkup-funnel-implementation-spec.md and outputs/checkup-funnel-prototype.html (also persisted as desktop artifact ainsure-checkup-funnel-prototype)
  • notes: Pushed back on the proposed sequencing — three of its four steps are mis-ordered for SMB commercial. Signup precedes value; carrier auth is the highest-friction door (Canopy’s own ~70% completion from credentials submitted; SMB owners often have no portal login because the incumbent agent owns the relationship; one Canopy Pull = one carrier, so a 4-carrier SMB is 4 MFA flows); and a serial pull is minutes of dead air. Replaced with identify → provisional picture → multi-rail intake (dec page > COI > renewal email > carrier login) → progressive analysis. Verified that ACORD 25 carries no premium, no deductibles and no exclusions — highest availability, lowest information, so it is a qualifier and renewal-date capture only. Analysis engine specced on a Tier A/B/C confidence model; Canopy returns PolicyForm[] with is_exclusion_endorsement and an ~85-value endorsement_type already normalised, removing the need for an ISO lookup table on day one. Findings deliberately lead with structural/administrative items (named-insured mismatch, misaligned renewal dates) that need no price competition — H-INS-9-aware and the concrete form of AX-INS-7. Central reframe: per H-INS-3 most check-up completers are mid-term and cannot convert, so the funnel’s primary output is a dated renewal pipeline, not a customer — Canopy’s Monitoring API supplies change detection for the credential-door cohort. Six-week build plan, instrumentation tied to the AX-INS-2/AX-INS-8 kill-gates, and compliance boundaries (credentials never touch our servers; analysis is ours, quote/recommendation happen under the partner’s license — counsel memo required before launch). Axiom check clean; flagged watch-item that door 5 (“we ask your agent”) and deep partner-AMS integration both edge toward AX-INS-4 territory.

[2026-08-16] query | technology vision & architecture for the 20-job product outline

  • source: Saar’s A product.md (20-job brokerage product outline) + request for tech vision/GTM and engineering answers. Research: two parallel passes on (a) AI policy comparison state of the art, (b) browser-agent portal automation and carrier appetite data. Load-bearing claims verified directly: Vals Legal AI Report task table, Coterie API docs.
  • new: insurance-technology-vision
  • updated: index; deliverables outputs/technology-vision-and-architecture.md and outputs/architecture-map.html (desktop artifact ainsure-system-architecture)
  • notes: Five-plane architecture (Acquisition / Placement / Service / Ledger&Compliance / the Brain) placing all 20 jobs with P0–P3 phasing, per-job engineering answers, build-vs-buy table, team shape, risks. Three governing rules: buy every rail own every representation; deterministic first, AI on the tail; the four compounding assets are the product (observed-appetite graph, form corpus, conversation corpus, renewal graph). Key evidence finding that inverted two prior recommendations: the Vals Legal AI Report shows AI beating lawyers on Document Q&A (94.8% vs 70.1%) but LOSING on redlining (65.0% vs 79.7%) — two-document differential with materiality judgment, the closest analogue to policy comparison, and the only task humans win; corroborated by CLAUSE (omission detection “exceptionally challenging”, citation matching <14%), ContractEval (best F1 0.644), and the Quebec insurance RAG benchmark (Gemini-2.5-Pro 74.65%→14.06% with naive retrieval). Applied Epic’s shipped policy checking is a paragraph text diff with materiality 100% human; no vendor publishes comparison accuracy at all. Design answer: recall-optimised, mandatory verified source spans (string-matched back or dropped), per-coverage-element measurement, silent-miss release gate. Portal automation recommendation reversed from build-an-agent to record-and-replay + LLM repair on the top 15–25 portals only (WebBench write tasks 46.6% vs WebVoyager 85.8% for the same agent; 60–120 actions per submission needs >99.5% per-action). Legal constraint is appointment termination, not litigation — strategic inversion is carrier-sanctioned signed-agent access via Web Bot Auth. Appetite reframed as a data-acquisition problem: Coterie’s own docs verified as stating the API has no appetite endpoint and the Integration Manager emails a spreadsheet, while requiring an AKHash (NAICS + description MD5). Two reclassifications of the outline: renewal prep is P1 not late (the renewal graph is the acquisition engine), COIs are an acquisition/retention channel not back-office; jobs 19–20 are axes not jobs. AX-INS-4 tension flagged explicitly — jobs 7–17 are the broker back-office FurtherAI/Qumis sell, and the only thing keeping us inside the axiom is that we build them for ourselves and never sell them.

[2026-08-16] ingest | meeting: 2026-08-16-directions-broker-jobs-to-be-done-map

  • source: raw/meetings/notetaker/2026-08-16-saar-nizan-11-05.json
  • new: 2026-08-16-directions-broker-jobs-to-be-done-map, broker-jobs-to-be-done-map, coverforce
  • updated: smb-insurance-portfolio-brokerage, us-insurance-distribution-economics, user-discovery-cadence, action-items, lemonade, saar-arbel, nizan-shifman, index
  • notes: Saar/Nizan session decomposing the AI-native broker into nine sequential jobs-to-be-done (qualification/exposure → commission reconciliation) as the spine for both V1 product scope and the deck’s Solution chapter. First appearance of the working name “Actually” (supersedes “Ainsure” from the 2026-08-10 deck sessions; final name/domain still deferred). CoverForce filed as the second remaining independent neutral quoting rail alongside herald.
  • ⚠ This log entry was reconstructed on 2026-08-16 by a parallel Cowork session. That session force-wrote wiki/index.md and wiki/log.md from a snapshot taken before this meeting was ingested, overwriting the original entry and index lines. Page content itself was never affected. The index lines and this entry were rebuilt from the page frontmatter; if the original log wording contained detail not derivable from the pages, it is lost. Root cause: concurrent sessions editing shared index/log files without re-reading first.

[2026-08-17] query | company research: Kanopi + Axle funding/what-they-do

  • source: web research (kanopicover.com, insuranceNEWS.com.au, Coverager, axle.insure, PR Newswire) — no raw source captured
  • new: kanopi
  • updated: axle, index
  • notes: User asked what Kanopi (kanopicover.com) and Axle (axle.insure) do and how much they raised. Kanopi filed as a new company stub: Australian ex-Expense-Check insurer/MGA/broker tooling platform, ~AU4M Gradient seed → 21.5M total) and 2026-08-17 homepage claims (220M+ losses recovered — all company-claimed); personal-lines-only caveat unchanged.

<<<<<<< HEAD

[2026-08-17] query | Canopy Connect intended use & customer consent scope

  • source: direct product/legal analysis from guy-barkat — no raw meeting or article; working from Canopy’s published Innovator Terms (last updated 2026-08-12)
  • new: canopy-connect-consent-scope
  • updated: canopy-connect, insurance-checkup-funnel-design, index
  • notes: Fixes the intended Canopy data flow (policy data + KYC → needs assessment → quoting → comparison → recommendation) as consistent with Canopy’s Innovator Terms, and records a recommended short-form consent above “Connect your insurance.” Draws a hard boundary against reading the consent as covering unrelated secondary uses (data sale/licensing, third-party database enrichment, unrelated model training) — those need separate consent. Flags Canopy’s “Licensed Third-Party Data” as a more-restricted category distinct from carrier-sourced policy data. Explicitly non-legal-advice; partially answers, but does not resolve, insurance-checkup-funnel-design’s open per-state solicitation-line question — counsel review still required before launch copy. =======

[2026-08-17] ingest | meeting: 2026-08-17-directions-greg-ehly-independent-broker-interview

[2026-08-17] ingest | meeting: 2026-08-17-directions-jim-coronado-multiline-broker-interview

[2026-08-17] ingest | meeting: 2026-08-17-directions-insurgrid-vendor-call

[2026-08-17] refactor | index alphabetisation + Notetaker fetch-path deviation

  • reason: (a) index.md category lists had drifted out of alphabetical order, which CLAUDE.md Section F step 7 requires; entries were re-sorted by slug while adding the day’s pages. (b) The sanctioned fetch_notetaker_meetings.py path could not run in this session — .notetaker.env is absent from the working copy and the environment has no network egress to api.monday.com — so the three meetings were retrieved through the monday Notetaker MCP instead, and the raw JSON files were not written.
  • touched: index, plus every page listed in the three ingest entries above.
  • notes: no content was lost in the re-sort (verified by set comparison against the previous version). Wiki resource: fields point at the canonical filenames the fetcher will produce, so a backfill drops straight in — the raw layer stays lossless and in the fetcher’s own schema rather than carrying a hand-transcribed copy. Backfill command: python3 skills/fetch-notetaker-meetings/scripts/fetch_notetaker_meetings.py --ids 6cdc7e05-20cd-48cb-9f76-d5b23481718d,bb6c9b65-a9eb-4cc3-be29-fe761b48a443,42e011ed-4b8d-4828-92f5-471ed5dd87da --save
  • core-team gate: Greg and Jim sessions carried all three founder emails in participants and passed cleanly. The InsurGrid call returned empty participant metadata, but the transcript names Guy, Nizan and Saar explicitly (Guy introduces the three by role) — treated as an override consistent with the precedent set repeatedly in state/notetaker/processed.md.

[2026-08-17] lint | post-ingest full check

  • ran: python3 scripts/okf_profile_lint.py — passed, 219 concept pages.
  • index integrity: 219 concept pages / 219 index entries, no duplicates, no dangling entries, no missing pages. No [[wikilink]] inside any YAML frontmatter. action-items.md has all four H2 sections in order; no item appears in both an open and a closed section; every action-item source meeting resolves.
  • flagged (all pre-existing, none introduced by today’s ingest):
  • not flagged but worth a decision next session: greg-ehly, jim-coronado, manya-khanna, vertafore, levitate, travelers and ezlynx all sit at source_count: 1 — single-source pages by construction, not orphans.

2d32828 (Ingest 2026-08-17 meetings: Greg Ehly, Jim Coronado, Insurgrid vendor call)

[2026-08-18] ingest | meeting: 2026-08-18-directions-moshe-tamir-gtm-and-underwriting-advisory

  • source: raw/meetings/notetaker/2026-08-18-moshe-guy-12-00.json (raw JSON backfill pending — same fetch-path deviation as 2026-08-17; see below)
  • new: 2026-08-18-directions-moshe-tamir-gtm-and-underwriting-advisory, moshe-tamir, danny-tsiddon, viola-fintech, insurify, white-label-distribution-gtm, underwriting-specialisation-required-for-durable-price-advantage
  • updated: insurance-distribution-aggregator-layer, insurance-seed-fundraising, smb-insurance-portfolio-brokerage, broker-jobs-to-be-done-map, policy-ingestion-requires-a-data-rail, guy-barkat, nizan-shifman, saar-arbel, index, action-items
  • notes: hour-long advisory call with a former Migdal deputy CEO / Generali global innovation lead, all three founders present (participants carried all three core emails plus moshe@axell-hub.com — gate passed cleanly). Substantively the strongest external challenge the direction has taken. Core claim: CAC is the binding constraint, not technology or insurance. Three recommendations follow — don’t start D2C (sell the tech to brokers on revenue share first), white-label to large distribution owners, and consider learning in Israel before the US. Separately, the sharpest strategic objection so far: a generalist broker cannot hold a durable price advantage without underwriting specialization or data enrichment, because carriers reprice or disengage from an unselected book.
  • axiom tensions surfaced, none amended (invariant 6 — axioms change only by logged team decision): AX-INS-4/AX-INS-8 (the broker-tooling path, argued this time as a learning step rather than as the business), AX-INS-9 (US-first, challenged on learning cost rather than substrate), AX-INS-7 (generalist positioning, challenged on carrier-side rate pressure — an argument the axiom’s own revisit trigger does not anticipate), and AX-INS-3 brushed but not violated (buying a broker for licenses and appointments, not for the book). Filed as a Team action item for an explicit decision.
  • countervailing note recorded on the new hypothesis: the underwriting objection sits in direct tension with the 2026-08-17 operator interviews — greg-ehly reported no minimums or ongoing conditions across ~15 appointments, and neither broker described a carrier repricing an agency over book composition. Moshe’s experience is also predominantly carrier-side and predominantly non-US.
  • flagged: two Hebrew-transcript company names are uncertain (“דאורו AI”, “קיסטה”) and are recorded as unconfirmed rather than researched. The consultancy name on moshe-tamir is inferred from his email domain, not stated. danny-tsiddon is a transliteration — confirm before outreach.

[2026-08-18] correction | retracting the 2026-08-17 transcript-truncation flag

  • reason: the 2026-08-17 ingest flagged the Jim Coronado transcript as possibly truncated at ~21 of ~41 minutes, inferred from the recording_duration field. The user backfilled the raw JSON on 2026-08-18; it ends at the same 21.1-minute mark with 111 segments. The MCP capture was complete and nothing was missing. recording_duration is not wall-clock speech time and should not be used to infer transcript completeness.
  • touched: log (original flag struck through with the retraction), state/notetaker/processed.md (all three 2026-08-17 rows updated from “BACKFILL PENDING” to backfilled, with the truncation note retracted).
  • notes: the three 2026-08-17 raw JSONs are now present and canonical, so those meeting pages’ resource: paths resolve to real files. The 2026-08-18 session is in the same pending state — .notetaker.env is still absent from the working copy and this environment has no network egress to api.monday.com. Backfill command: python3 skills/fetch-notetaker-meetings/scripts/fetch_notetaker_meetings.py --ids bc2a9f9e-e166-4217-b75a-93de8c090d48 --save

[2026-08-18] lint | post-ingest full check

  • ran: python3 scripts/okf_profile_lint.py — passed, 230 concept pages (up from 219).
  • index integrity: 230 pages / 230 entries, no duplicates, no dangling entries, no unindexed pages. No frontmatter wikilinks. Every new page has inbound links (7–15 each), so nothing landed orphaned. All action-items.md source wikilinks resolve.
  • flagged (unchanged from 2026-08-17, all pre-existing): broken [[pagaya]] and [[maccabi-netanya]]; kanopi still holds [[meeting-slug]]/[[source-slug]] placeholders and source_count: 0; insurance-checkup-funnel-design still has no ## Evidence section.

[2026-08-19] query | Coverdash refresh — embedded-distribution evidence check

  • source: web research pass (PRNewswire 2026-08-12; coverdash.com partnership archive; TechCrunch/Yahoo Finance 2024-02-28; BusinessWire + fintech.global; Tracxn / PitchBook / CB Insights / Crunchbase)
  • updated: coverdash, smb-insurance-competitive-landscape, smb-insurance-portfolio-brokerage, wiki/index.md
  • notes: Triggered by the 2026-08-19 session comparing three directions (D2C brokerage / platform-for-brokers / enterprise-aggregator revshare), in which Coverdash is the load-bearing counter-evidence for the whole embedded/white-label family. Two corrections to the 2026-08-01 picture: (1) “mostly long-tail partners, no tier-1 platform wins” is no longer true — Coverdash embedded into Housecall Pro on 2026-08-12 (200,000+ home-service pros, “1 in 4 American homes”), and had already signed LendingTree (Nov 2024), Bizee (1M+ LLCs) and Swyft Filings; full dated partner list now on the company page. (2) The traction figures in circulation (100+ embedded partners, 30% MoM customer growth, 5x headcount) are Ralph Betesh’s February 2024 Series A claims, not current, and he declined revenue specifics at the time. Unchanged and now better sourced: still $16M total, still no round since 2024-02-28 (four independent trackers agree), and still zero disclosed attach rates anywhere in SMB embedded insurance. Net read: scaled distribution partners are demonstrably winnable; what remains unproven is that they convert into volume — the gap between a 200K-user partner and a 2.5-year funding silence is now the informative fact, replacing the old “they can only sign small partners” reading.
  • flagged: H-INS-6 (“Embedded distribution is an experiment, not a plan”) cites coverdash as evidence. Its attach-rate clause still holds; its implicit “no tier-1 partner” support does not. NOT amended here — heuristics change only by explicit team decision. Recommend the team review it alongside the direction discussion.

[2026-08-19] query | Harper GTM refresh — wedge, unit economics, and the abandoned broker-tooling plan

  • source: web research pass (harperinsure.com 2026-08-19; TechCrunch 2026-02-25; Emergence Capital thesis post; Antler founder story; YC company page)
  • updated: harper, wiki/index.md
  • notes: Asked in-session what Harper’s GTM actually is. Four corrections to the existing picture. (1) The wedge is hard-to-place risk, not moment-of-new-risk — the site sells to businesses “told your business is too tricky to insure, or too late” and firms “other insurers turn down,” across ten main-street verticals (construction, restaurants, auto sales, garage services, childcare, adult care, manufacturing, hospitality, security, transport). The 2026-08-10-directions-huri-pitch-deck-build reading of “reach businesses exactly when they hire or expand” is not supported by their public positioning. (2) “1,000+ new customers/month” is a throughput/capacity claim, not an achieved acquisition rate — actual is ~5,000 customers in 13 months (385/month). (3) **Unit economics: 5,000+ businesses / just over 1,200 average premium per account (900K of commission revenue against $47M raised. (4) Site now claims a 98.78% renewal rate — the first published retention figure in the wave, unaudited and over a book barely one cycle old.
  • flagged: H-INS-1 (“thin accounts kill brokerages”; floor ≥3K commission) is now in direct, testable conflict with the best-funded competitor’s actual book. Either Harper is building Verifly’s corpse with better tooling, or AI throughput moves the floor and H-INS-1 needs a revised threshold. NOT amended — heuristics change only by explicit team decision. This belongs on the agenda before the ~2026-08-27 alon-huri follow-up. Also relevant to the same discussion: Harper originally planned to sell an outsourced sales service to existing agencies and abandoned it because brokerages resisted it as a threat to their producers — the category’s clearest empirical evidence on the sell-to-brokers path, bearing directly on the three-way direction debate held this session.

[2026-08-20] ingest | meetings: 2026-08-19 ×2 and 2026-08-20 ×2

[2026-08-20] correction | mistyped meeting ID in the 2026-08-20 ingest

  • The 19 Aug “Directions meeting” ID was first written as c1a0cf73-4318-4f6f-8f24-ec1a3a0cf73 in state/notetaker/processed.md and in the backfill command above. That string is wrong — a transcription slip, not a source error. Verified against the Notetaker index and corrected in both files to c1a0cf73-43b9-4418-8f6f-dea896180475. The wiki content was unaffected; only the dedupe key and the backfill command were, and both would have silently failed.

[2026-08-20] lint | post-ingest full check

[2026-08-21] refactor | Clear stale Notetaker backfill flag

  • reason: state/notetaker/processed.md still marked the 2026-08-18 Moshe Tamir meeting as “RAW JSON BACKFILL PENDING”, but the raw JSON was in fact backfilled and committed in 733cb8b.
  • touched: state/notetaker/processed.md (no wiki concept pages changed)
  • notes: Verified raw/meetings/notetaker/2026-08-18-moshe-guy-12-00.json is complete — 470 transcript segments, summary, 7 topics, 4 action items, and all four participants (guybarkat2, saararbel1, nizanshifman, moshe@axell-hub.com). No re-fetch needed. Separately, the weekly sync could not run this session: .notetaker.env still contains the literal placeholder key from the skill docs, so monday returns HTTP 401. Meetings after 2026-08-18 remain unchecked until a real API key is in place.

Graph View

Backlinks