A three-step onboarding playbook for seeding aswritten.ai from historical recordings, articulated around the enterprise customer's ARB transcripts (May 2026). [Step 1: code-base review] (Claim_CodeBaseReviewStep1) runs Claude/Copilot CLI over the code with structured questions, persisting a paragraph per workflow to memory. [Step 2: topic analysis] (Claim_TopicAnalysisStep2) surfaces what a corpus covers and its open questions, then scopes memories — but [requires a domain expert] (Claim_UserDomainCommentaryRequired) (at the enterprise customer, only the customer's chief architect) to annotate, or seeding goes sparse or wrong. [Step 3: ongoing iteration] (Claim_OngoingIterationStep3) treats each new transcript as a diagnostic, with a 30-minute voice memo filling gaps. The pattern is pitched as an Enterprise playbook that [generalizes beyond the enterprise customer] (Claim_OnboardingMethodologyGeneralizes) without new tooling, while [Step 2 itself] (Claim_TopicAnalysisPlatformFeature) is flagged as a future Platform feature. Note a dated reframing: a [May 21 claim validates bulk AI summarization] (Claim_BulkIngestionViaAISummarizationValid) — what the customer's chief architect did — as effective, retracting the earlier 'failure' framing, and [ranks three ingestion methods] (Claim_ThreeRankedIngestionMethods) with a reconciliation-based hybrid as best.
Witness phrases
verbatim read the code base, understand what the core pre-built workflows are and write a paragraph about each one and persist that paragraph to collective memory
verbatim The transcript is the diagnostic, not the memory.
verbatim Dump the transcript into aswritten and over the course of a couple weeks or even a couple months, you'll converge on having pretty much everything covered.
verbatim The 'we discovered it by failing' framing was a bad AI artifact; they did not fail.
verbatim Method 3 is best but depends on the reconciliation tool, which is not yet built.
verbatim Without that, the seeding will produce sparse or wrong knowledge.
verbatim Dated reversal: May 1 framing treats direct historical ingestion as an anti-pattern to avoid (and possibly block via UX), but May 21 declares bulk AI summarization valid and effective, superseding Foundation Claim_DirectIngestionIsTrap.
verbatim The stated-best ingestion method (batched + human reconciliation pass) is aspirational — it depends on a reconciliation tool that does not yet exist; only methods 1 and 2 are actually available.
verbatim The whole methodology is gated on a hard precondition: a customer domain expert who can validate the topic analysis. At the enterprise customer that is a single person (the customer's chief architect); without such a reviewer the pattern degrades to sparse or wrong knowledge.
verbatim Tension between 'no new tooling needed, just sequence existing capabilities' (generalizable playbook) and the claims that Step 2 should become a Platform feature and that the best method needs an unbuilt reconciliation tool.
In this area
ClaimClaim_BulkIngestionViaAISummarizationValid Bulk historical ingestion via AI summarization is a valid and effective onboarding method. Scarlet explicitly advised the customer's chief architect to use it for the enterprise customer's ARB transcripts because it would be faster than pro
ClaimClaim_CodeBaseReviewStep1 Step 1 of the three-part seeding pattern: use Claude / Copilot CLI connected to aswritten.ai to read the code base with structured questions ('read the code base, understand what the core pre-built workflows are and write a paragraph about
ClaimClaim_OnboardingMethodologyGeneralizes The three-part seeding methodology generalizes beyond the enterprise customer: any domain expertise codified in historical recordings (sales calls, customer interviews, OKR retrospectives) is a candidate for the same pattern. This is a stat
ClaimClaim_OngoingIterationStep3 Step 3 of the three-part seeding pattern: as ARB meetings happen, ingest each transcript with active review. The transcript is the diagnostic, not the memory. When the system fails to answer a question well, the user does a 30-minute voice
ClaimClaim_ThreeRankedIngestionMethods Three bulk ingestion methods ranked: (1) Full summarization across the set — what the customer's chief architect did, effective; (2) Per-transcript memory-by-memory ingestion — slow and overwhelming; (3) Target method — batched ingestion fo
ClaimClaim_TopicAnalysisPlatformFeature The topic analysis step (Step 2) is currently entirely operator-driven and is a high-value activity that could become a Platform feature. The anti-pattern (direct historical ingestion) should be reflected in onboarding documentation and pos
ClaimClaim_TopicAnalysisStep2 Step 2 of the three-part seeding pattern: rather than ingesting six months of ARB transcripts directly, use Claude to perform a topic analysis across the corpus — 'what are these talking about? what are the open questions? what are frequent
ClaimClaim_UserDomainCommentaryRequired The methodology assumes the customer has domain experts who can review and annotate the topic analysis output. The user — not aswritten.ai staff — provides the architectural commentary at Step 2. Without that, the seeding will produce spars