---
name: mc-persona-capture
description: >
  Author a deep persona capture (philosophy, style incl. spoken delivery,
  Q→A exemplars, eval probes, self-critique) for one named person into
  captures/<slug>/. FRONTIER-MODEL WORK: run only with the strongest model
  available (Fable-class; Opus acceptable fallback) — the output quality IS
  the product and cannot be redone cheaply. Use when the user says "capture
  <name>", "author the <name> persona", or "next capture". One persona per
  session; suggest /clear before starting the next.
---

# MentorCore Persona Capture (frontier authoring pass)

Author the checked-in distillation tier for one persona. Full spec:
`docs/PERSONA-CAPTURE.md` — read it first if not already in context. Skeleton:
`captures/_template/`.

## Step 0: Gate on model tier

This skill exists to spend scarce frontier-model capacity on synthesis a weaker
model cannot redo. If you are not a frontier-tier model (Fable-class, or Opus as
fallback), STOP and tell the user which model to switch to. Do not "do your
best" — a mediocre capture checked into git is worse than none.

## Step 1: Setup

1. Resolve the slug (kebab-case). Create `captures/<slug>/` from
   `captures/_template/`.
2. Resolve the **capture focus** (spec: PERSONA-CAPTURE.md §Focused captures):
   take it from the arguments if given (`<name> — focus: ...`), else check the
   persona's roster line in `docs/BUILDPLAN.md` Phase 1e for a `focus:`
   annotation. If a focus exists, it steers the depth budget, the exemplar
   register (e.g. coaching dialogues, not interviews), and the eval probes —
   but the whole-life sections (core worldview, evolution, anti-beliefs,
   blind spots) are still mandatory. Record it in `capture.yaml:
   capture_focus`. No focus → whole-life capture, `capture_focus: null`.
3. Fill `capture.yaml` last (after authoring, so `confidence` and `notes` are
   honest). `captured_by` = your exact model ID.

## Step 2: Author, in this order

Order matters — later files reuse the loaded context from earlier ones.

1. **`philosophy.md`** (2,000–5,000 words; all 8 spec sections). Flag
   uncertain claims inline with `<!-- verify -->`. Depth over breadth.
2. **`style.md`** (1,000–2,500 words; all 8 sections including **Spoken
   delivery** — write it as punctuation-as-pacing instructions the runtime
   LLM can follow; there is no SSML downstream). The Do/Don't pairs are the
   highest-value section: ≥5 pairs, same idea in-voice vs off-voice.
3. **`qa-exemplars.jsonl`** — 20–40 pairs, `synthetic: true` on every line.
   Cover: core domains, transfer questions (mental models on topics they never
   addressed), hostile/skeptical questions, and era/ignorance triggers.
4. **`quotes.jsonl`** — ≥50 quotes if the corpus supports it, verbatim text
   only, spread across topics. Provenance honesty is a hard rule: mark
   `verified` ONLY if you actually fetched and checked the source this
   session and recorded `verify_url`; otherwise `model-memory` with a
   suggested `verify_url`. Never fabricate URLs. The cheap `/mc-quote-verify`
   pass upgrades these later — quarantine is safe, fake verification is not.
5. **`eval-questions.jsonl`** — 15–25 probes where THIS persona is most
   likely to break (`era`, `misattribution-bait`, `core-belief`,
   `style-stress`, `anti-belief`, `transfer`).
6. **`sources.md`** — what informed the capture + what the pipeline should
   ingest later (with fetch hints).
7. **`voice.yaml`** — policy + `pacing_notes` (your judgment, write it now)
   and any reference URLs you are confident exist. Leave `clean_segments`
   empty rather than inventing timestamps — `/mc-voice-refs` fills them.

## Step 3: Self-critique pass (do not skip)

Re-read the whole capture as a hostile expert — a lifelong scholar of this
person. Ask: What is wrong? What is missing? What is flanderized (the
caricature version of them)? Which era is over/under-represented? Revise the
files, then record residual gaps honestly in `capture.yaml: notes` and set
`confidence` accordingly.

## Step 4: Commit and hand off

1. Tick the persona's checkbox in `docs/BUILDPLAN.md` Phase 1e.
2. One commit for the whole capture: `Capture: <Display Name> (frontier authoring pass)`.
3. Tell the user: remaining roster per BUILDPLAN, and that quote verification
   (`/mc-quote-verify <slug>`) and voice segments (`/mc-voice-refs <slug>`)
   are cheap-model follow-ups — do NOT spend frontier time on them.
4. If context is running long, suggest `/clear` before the next capture.

## Budget degradation (if frontier time is running out)

Cut from the bottom, never the top: philosophy.md > style.md > exemplars >
self-critique > eval-questions > quote curation > sources > voice refs.
A complete philosophy+style for fewer personas beats partial captures for many.