Synthesis Meeting Prep

Meeting Prep

Version 1.0.0 (2026-09-20): first release — the content contract for meeting preparation (requirements R1–R12), the 64-factor inventory, six structure variants, reader profiles, the debrief half, and a mechanical prep linter (scripts/prep_lint.py).

An agent preparing a principal for a meeting is not a summarizer. It is the principal's chief of staff for one room: it figures out what the room needs, what the principal wants, what could go wrong, and puts exactly that on one scannable page. The content contract lives in references/requirements.md — R1–R12 are normative. This file is the operating protocol.

1. Configuration contract

All person-specific state lives outside the skill, in private paths the skill reads at load time:

~/.synthesis/meeting-prep/principal.json     # role, goals, authority, tells
~/.synthesis/meeting-prep/readers/<id>.md    # per-person reader profiles (R12)

principal.json holds what the skill must know about the principal: current role and org, professional and personal goals, what they can commit in a room, their known positions, and their tells under pressure. Reader profiles hold relationship, technical depth, what the person cares about, what they have already been told, and what landed last time. Both are scaffolded by scripts/prep_init.py (init once, add-reader per regular) and finished by interview the first time the skill runs for a principal — never copied from another machine's private files. Missing config degrades gracefully: the skill asks the three questions it cannot proceed without (who is the reader to you, how technical, what is the meeting's job) and drafts anyway.

2. The prep loop

For every meeting, in order:

1. Gather. Pull the knowledge factors (inventory §4): workspace OKF base, prior transcripts with these participants, comms since, tickets/commits/deliverables, docs and decks, reader profiles, 2–3 samples of the principal's own writing to a comparable reader (R13 — required; if none exists, say so in the pack), external research for new faces, and re-verified numbers. Record the basis (R10) as you go — source, date, what it contributed.
2. Model. Write the audience model first (R2): each key participant's relationship, technical depth, and cares; the power map; the disclosure boundary. Then weigh the factor inventory and name the drivers (§6, factor 63) — the 3–5 factors that set this pack's shape. If a factor is unknown and material, say so; never fill a material gap with a guess presented as fact (factor 62).
3. Size. Duration, format, prep budget, and energy set the length (factors 4, 31, 32, 57). State the sizing in one line on the pack.
4. Draft. Pick the structure variant (structures.md), drop inapplicable sections, and write to R1–R9: discuss-forward, dense, translated accomplishments, strategic altitude, specific credit, grounded register, scannable, positive framing. Mark confidence inline — fact, inference, guess.
5. Lint. Run scripts/prep_lint.py on the draft. It checks the mechanical half of the contract: footer placement, register leaks (ticket numbers and infrastructure terms for non-technical readers), empty-sentence patterns, name-drop openers, unverified numbers, missing basis. Fix every finding or record why it is wrong — the linter is a backstop, not the author.
6. Deliver. The pack plus its capture half, filed where the principal's workflow expects it (workspace meeting-preps/ by default, <date>-<slug>-prep.md), and the two-line drivers note in chat. Never deliver brackets for the principal to fill in — assembled packages, not assembly kits.

3. The debrief loop

After the meeting, when a transcript or notes exist:

1. Verify the record. Transcript-primary: the verbatim record is the only source for attribution-bearing claims. Never derive who said or decided what from a summary.
2. Extract. Decisions (verbatim where it matters), commitments with owner and date — no ownerless commits — quotes worth keeping, and the prep-vs-reality gap: what the prep missed or mis-weighted.
3. Draft the follow-ups. Assembled, in the principal's register, ready to send — not a list of "follow up on X."
4. Update the profiles. Fold the gap into the reader profiles: what landed, what didn't, what they were told. This is how prep compounds.
5. Capture durable facts. Anything that belongs in the workspace knowledge base goes there now, with provenance — not "later."

4. Integration seams

• Daily rituals (synthesis-daily-rituals lead-time prep packs) own the scheduling machinery — which meetings need packs by when. This skill is the content contract that machinery drafts against.
• Meeting transcripts (synthesis-meeting-transcripts) own transcript fetch, naming, and the transcript-primary rule the debrief depends on.
• OKF (synthesis-okf) is the workspace knowledge interface for factor 37.
• Console renders packs filed under meeting-preps/; the basis statement is what makes them trustworthy there.

5. What the skill refuses

• Prep for a reader it has not modeled — no model, no draft.
• Numbers it has not re-verified — stale figures kill credibility.
• Guesses presented as facts — mark confidence or cut the line.
• Prep that contradicts the principal's recorded positions.
• Anything that crosses the disclosure boundary for the room's least-trusted attendee.