Skill explainer + doc index

Transcript-Ingest V3

A nine-pass-plus-prime protocol that turns any meeting transcript into a routed set of vault artefacts (meeting note, decisions, project updates, considerations, pain points, tasks) with a Confirm gate (reviewer reassignable) before any writes happen.

V3 . ready vault-internal skill validated 2026-05-26 corpus engine queued

What it is

A skill that solves the "transcripts are the OS's primary input signal but extraction is variable" problem. Before V1 we did freeform read-and-summarise; that lost the load-bearing chain in the 23 May Strategy Session 2 ingest. The skill replaces freeform with a strict ten-pass protocol that extracts, routes, gates on your confirmation, then applies writes across the vault.

The skill is a vault-internal protocol document, not a packaged plugin skill. It lives at Skills/transcript-ingest/SKILL.md and fires via a verbal trigger registered in root CLAUDE.md § Shortcuts.

Trigger phrases that fire the skill: intake this transcript, process this transcript, ingest this transcript, fan this out. Also auto-fires when a Fireflies transcript is fetched via MCP with no other instruction.

How it works . the ten-pass protocol

Pass 0 primes context from the vault. Passes 1 through 6 extract. Pass 7 produces a routing ledger. Pass 8 is the Confirm gate. Pass 9 applies the writes.

0
Context prime

Read-only warm cache: active workstreams (root CLAUDE.md line 7), team roster, project list, last 30 decisions, last 10 meetings. Prevents asking you about entities the vault already knows. Informational only . canon trumps it, transcript trumps both.

1
Entity extraction

Every named person, company, product, place, tool, concept. Role-context disambiguation resolves "Verner" to Werner Snyman inline without flagging. New entities flagged for your confirm.

2
Decision extraction

Every "let's decide X", "we'll go with X", "I want X". Each decision: what's being decided, what it supersedes, what changes downstream, confidence.

2b
Decision-arc tracking

Follows each decision-shaped thread end to end. If you opened on A, debated B, landed on C, Pass 2b writes only C as a decision. Rejected alternatives become considerations (Pass 5b). Catches debates that V2 would have logged as two competing decisions.

3
Framing extraction

Definitional concepts and concepts returned to more than once. The pass that catches load-bearing chains: if two or more framings share linkage, Pass 6 produces a chain summary.

4
Action + project-state extraction

Owner-action pairs (tasks) plus status updates, blockers, milestone shifts, capacity claims (project-state items). Distinguishes "Werner will ship Friday" (task) from "Werner is blocked on Tony" (project-state).

5
Open question extraction

Things the speaker doesn't know. Actionable questions become research stubs; strategic questions stay as Pending in the meeting note.

5b
Considerations extraction

Options being weighed without resolution. Different from open questions (things unknown). Light considerations stay in-note; substantial ones promote to Intelligence/considerations/.

5c
Context-update extraction

Statements that update static-ish context files: pain points, competitor intel, market intel, strategy updates, team/stakeholder updates, services updates, brand updates.

6
Orchestrator pass

Re-read with Passes 1-5c in view. Catches things assigned to two passes, chains linking multiple decisions, contradictions, residue not captured by any pass. Produces the chain summary callout.

7
Multi-target routing ledger

Single table with one row per artefact or append the skill proposes to write. Columns: topic, source pass, category, route path, form, owner, your_move. The contract presented to you.

8
Confirm gate

The ledger is rendered in chat. Nothing writes until you reply apply, ship, go, or yes, or amend specific rows. Silence is not consent. kill aborts.

9
Apply

Sync probe (Rule 15 V20) if 3+ Relay-shared rows. Meeting note first as source-of-record. Each derivative artefact in ledger order, wikilinked back. Tasks last. Dual-write per Rule 15 V21. Em-dash sweep. Report listing every path.

How to invoke

Three ways to fire the skill, in order of how often you'll use them.

1. Verbal trigger in any session

Paste a transcript into chat, or upload a file, or just point at one already in Drive, then say one of the trigger phrases:

intake this transcript
process this transcript
ingest this transcript
fan this out

The skill runs Passes 0-6 in the session, presents the routing ledger from Pass 7, waits for your apply, then runs Pass 9.

2. Auto-fire from Fireflies fetch

When a Fireflies transcript is fetched via the MCP connector in-session and you haven't said what to do with it, the skill fires automatically. Same protocol, same gate, no manual trigger needed.

3. Drive Gemini Notes Doc fetch

Google Meet auto-generates "Notes by Gemini" Docs that contain both the Gemini summary AND the full verbatim transcript. Fetch by Drive ID or by title; same protocol applies. This is how the 22 May Fiorella meeting was ingested as the V3 validation run.

What's installed and usable right now

Ready to use

The skill itself (Skills/transcript-ingest/SKILL.md V3), the verbal trigger (root CLAUDE.md Shortcuts entry), dual-write discipline (Rule 15 V21 in canon), task scoring (Rule 20 in canon), and Pass 0 sources (all the vault paths Pass 0 reads exist and are populated).

Queued amplifiers

Back-catalog import of 100+ Gemini Notes Docs from Drive into Intelligence/meetings/_archive/. Cross-transcript corpus engine at Resources/engines/transcript-corpus/. Both tasks live in tom-king Tasks.md at low priority, gated on your go-ahead.

The plugin question

The skill is a vault-internal protocol document. It does NOT appear in Cowork's <available_skills> auto-discovery list. That means: invoked via verbal trigger, yes; auto-fired by Claude reading available skills, no. To get the latter, the skill would need to be packaged into the cw-house-v1.0.0 plugin or mirrored into Skills/external/ per canon Rule 29. That's deferred.

Validation status

V3 was validated on 2026-05-26 against the 22 May 2026 Tom-Fiorella Gemini Notes Doc. End-to-end exercise: Pass 0 primed 10 team profiles + 7 contractors + 23 projects + 30 decisions + 10 meetings; Pass 1 disambiguated five mis-transcribed names without flagging; Pass 2b resolved the subscriber-vs-investor priority arc into one decision record with a supersession trail; Pass 6 produced an AI-native-operating-mode chain summary tying five framings together; Pass 9 wrote six files and two context appends with em-dash sweep clean.

Two entity flags surfaced for your confirm: "Santi" (LATAM developer, not in roster) and "Farfa Jr" (probable Roy nickname, unresolved by Pass 0). Both queued as tasks in tom-king Tasks.md.

Doc index

Everything related to the skill, the framework, the validation run, and the queued amplifiers. Click any path to open in Obsidian.

The skill

DocPurposeStatus
Skills/transcript-ingest/SKILL.md The V3 skill itself. 10-pass protocol, hard rules, performance criteria. V3 active
Intelligence/research/2026-05-23-transcript-ingest-skill-brief.md Original May 23 brief that specced the protocol. Source of the 5+1 pass design. reference
CLAUDE.md § Shortcuts Root CLAUDE.md Shortcuts section. Registers the verbal triggers. live

Canon dependencies

DocPurposeStatus
.claude/rules/cloudworkz-canon.md Canon. Pass 1 checks every entity against this. Rule 15 V21 dual-write. Rule 20 task scoring. Rule 21 V33 session-close discipline. Rule 19 sync-recovery. V33 active
SKILL.md hard rules In-skill hard rules: Pass 0 informational not authoritative, Pass 6 mandatory, Pass 8 non-negotiable, dual-write Rule 15 V21, em-dash sweep at Pass 9. embedded

Validation run output (Fiorella 22 May 2026)

DocPurposeStatus
Intelligence/meetings/2026-05-22-tom-fiorella-investor-priority-shift-V1.md Source-of-record meeting note. Contains chain summary, entities, decisions list, Pass 2b arc trail, framings, project-state, considerations, tasks summary, entity flags, routing manifest. written
Intelligence/decisions/2026-05-22-investor-first-two-track-V1.md Decision: investor outreach is primary priority. Two-track strategy on the 30k hot list. Carries supersedes: for prior Sales-Loop framing. confirmed
Intelligence/decisions/2026-05-22-hot-list-email-campaign-V1.md Decision: first-of-kind email + retargeting campaign on the 12-year telephone-only hot list. Owner Fiorella. confirmed
Intelligence/decisions/2026-05-22-outcome-based-learning-V1.md Decision: training is outcome-based not curriculum-based across the team. Applied to Fiorella + Claudia + future onboarding. confirmed
Context/strategy.md Append: dated section on investor-first priority + two-track framing. appended
Context/pain-points.md Append: new "Operator / internal pain points" subsection (Tom-as-bottleneck + markdown-unfriendly). appended
Team/cloudworkz/Profiles/tom-king/task-list/Tasks.md Five new tasks under "Fiorella ingest follow-ups": Academy plan, screen recordings, investor-first cascade, two entity-flag confirmations. local only

Queued amplifiers (not yet built)

Doc / targetPurposeStatus
Intelligence/meetings/_archive/ (future) Bulk import of 100+ Gemini Notes Docs from Meet Recordings/ Drive folder. Independently shippable. Precondition for the corpus engine. Task: transcript-corpus-backcatalog in tom-king Tasks.md. queued
Resources/engines/transcript-corpus/ (future) Scheduled engine maintaining a registry of themes, decisions, framings, projects, people across the meeting corpus. Pass 0 + Pass 2b call into it for cross-transcript retrieval. Task: transcript-corpus-engine in tom-king Tasks.md. queued
cw-house-v1.0.0 plugin (future) Plugin packaging that would make the skill auto-discoverable in Cowork's <available_skills> list. Deferred until plugin scaffolding lands. deferred

Entity flags surfaced . awaiting your confirm

Two flags from the validation run that the skill could not auto-resolve. Both queued as tasks but listed here so you see them in one place.

V21 dual-write status

Of the seven files touched by the validation run, six are fully dual-written (vault_write to Relay + local Write/Edit, identical content, in that order). One exception: tom-king Tasks.md edits this session (V3-validate tick + new Fiorella section) are local-only; the rest of the file was vault-written earlier in the same session. Yjs CRDT will reconcile on the next Tasks.md write via MCP, or you can run the Rule 19 toggle (Settings → Community plugins → BenAI Relay off, three seconds, on) if you want immediate alignment.

What to do next

Three things you might pick, ordered cheapest first.