Bootstrap Persona

You are building the durable knowledge that lets a later agent write prose that belongs to an existing body of work — without reading that body of work again. The input is a description and some pointers. The output is a persona profile per conventions/personas.md: cited, measured, dated, and tested against a piece it was not allowed to see.

Read the convention first. It defines the layout, the modes, the excerpt rule, the staleness rule, and the boundary between the human-authored seed and this skill's derived output. This file is the regimen.

Operating principles

1. Read the work; do not infer the writer. Everything in the profile comes from text you actually read this session, cited by URL or path with an accessed date. No claim about how this author writes may rest on general knowledge of the venue, the genre, or the person's reputation — and if the corpus is thin, the profile says so rather than filling in.
2. Measure what can be measured. "Long, elegant sentences" is not a constraint; "median 24 words, range 6–61, with a short declarative closing most paragraphs" is. Count. State the sample the count came from. An approximate figure honestly labeled beats a confident adjective.
3. A profile is falsifiable or it is decoration. One or two pieces are held out before any reading begins and are used to test the profile's predictions (Phase 4). What the test reveals is written down — including where the profile was wrong.
4. Style is not only sentences. The reusable knowledge is as much structural and intellectual: how pieces open and close, how sections are grammared, what apparatus appears, how long pieces run, which thinkers recur, what has already been argued and where. A voice print alone produces pastiche with nothing to say.
5. Distill; never hoard. Notes and short excerpts, never wholesale copies of the corpus — the convention's excerpt rule (a critic's quotation, up to ~40 words, each with a stated purpose and source). Store the distillation, not the text.
6. Separate description from prescription. Observations belong in the derived profile; instructions about what a future draft should do belong in the seed, which is the human's document. When you believe a tic should be rationed, record the observed frequency and say so as an observation with a recommendation attached — do not silently convert taste into law.
7. The anti-signature matters as much as the signature. Imitation fails by over-production: the move the author makes twice per essay, an imitator makes nine times. Every signature move gets a measured rate, and the profile states what this author demonstrably never does.
8. Respect the mode. usage governs everything downstream. Under any mode but self or consented, the profile carries a prominent header stating what it may and may not be used for, and positions.md records only positions the author demonstrably took, quoted and cited.
9. Report the coverage honestly. Corpus size, how much was read, what was skipped and why, and the staleness verdict. A profile built from three posts is useful and must not pretend to be built from thirty.

Workspace

Everything is written under the persona directory — creating it if the persona is new, and migrating a legacy flat <id>.md into <id>/PERSONA.md if that is what you find, per the convention:

<personas_dir>/<persona-id>/
  PERSONA.md            # the seed — preserved; only appended to, with a marked note
  docs/                 # optional, operator-supplied corpus files: READ ONLY, never write here
  profile/
    index.md            # the hub other skills load: what exists, staleness, how to use
    voice-print.md
    structure-repertoire.md
    lexicon.md
    exemplars.md
    positions.md
    continuity.md
    audience-and-venue.md
    calibration.md
    provenance.md
    corpus/
      index.md
      <piece-slug>.md

Write each file as its phase completes; do not hold the profile in context and write it at the end. A corpus pass is long, and the per-piece notes are the expensive part — losing them to a compaction is the failure this layout prevents.

Execution model

Six phases. Phase 1 (the corpus read) is the heavyweight one and the natural place for parallelism.

• Where the harness supports subagents: after Phase 0, fan the corpus out — one subagent per piece or per small batch — each given the piece's URL or path, references/corpus-note.md, the excerpt rule, and the explicit instruction never to open the held-out piece (name it, by URL). Each writes its own profile/corpus/<piece-slug>.md. Then do Phases 2–5 yourself: synthesis needs one mind holding all the notes.
• Where it doesn't: read pieces sequentially, writing each note before opening the next, and never re-read a piece from memory in a later phase — read the note.

Do not synthesize from an archive listing, titles, or search snippets. A voice print derived from summaries is a plausible fabrication, which is worse than nothing because everything downstream will trust it.

Composing a composite persona (mode: compose)

When the seed declares Members, this is not a research task and the six phases below do not apply. A composite is a union of its members, never an average (the rules are in conventions/personas.md); its profile is derived from theirs, and no corpus is re-read.

1. Resolve the members. Each must exist and be a leaf persona — a composite of composites has no terminating routing rule; refuse it and say why. Note which members have profiles: a member with only a seed contributes its seed's claims, marked seed-only.
2. Settle mode and usage. usage is the most restrictive of the members' modes, and the composite may not grant a permission a member lacks. State the resolved mode and which member forced it.
3. Write profile/register-map.md — the routing rules a consuming skill will use to pick one member: per member, the venues, piece types, length ranges, and apparatus norms that indicate it, taken from each member's audience-and-venue.md and structure-repertoire.md. Add a default member and an explicit list of genuinely ambiguous cases. Routing rules that overlap silently are the failure mode here; make the overlap visible instead.
4. Union the substance. positions.md and continuity.md merged across members, every entry keeping its member, its piece, and its citation. Then do the work only a composite can: record cross-register relations — the same position argued twice in different registers, a position revised between them, a term used in both senses, an essay that popularizes a manuscript. Flag apparent contradictions between members as open questions for the human; do not resolve them yourself, and do not smooth them into agreement.
5. Optionally shared-lexicon.md: terms attested in all members, with each member's usage and any difference in sense. A term from one member only does not belong here.
6. Write no voice files. No voice-print.md, structure-repertoire.md, or exemplars.md at composite level — index.md points at the members' instead.
7. profile/index.md names the members, the register map's headline rules, the default member, the union counts (positions, continuity entries, cross-register relations, unresolved contradictions), and a staleness verdict equal to the worst of the members', with each member's profile date listed. provenance.md records which member profiles were read, their dates, and the composition date.
8. Calibrate the routing, not the voice (when calibration is on): take three to five real pieces spanning the members — identified by title, venue, and date only — and predict from the register map alone which member each belongs to. Record the misroutes and fix the map. A composite's failure mode is sending a piece to the wrong register, so that is what its calibration file measures.

Re-running compose after a member is bootstrapped or refreshed is cheap and expected: a composite is a view over its members, and a view that lags its sources is worse than no view.

Phase 0 — Resolve, consent, census

1. Resolve the persona. Locate the seed per the convention's discovery order, or create one from corpus plus whatever the operator has stated about the persona in this request (a seed is a short human document; write it as one, and mark it as agent-drafted for the human to correct). Under mode: refresh, read profile/provenance.md and profile/calibration.md first: reuse before recomputing.
2. Settle usage. From the seed's Mode, or the input, or default study with a prominent report. If the mode is consented, record who consented and how (one line, in provenance.md).
3. Take the census. Enumerate the corpus: every piece, with title, URL or docs/ path, publication date, and approximate length. An archive page usually gives this; a local corpus gives it from the files — and where a docs/ file's date or title is not recoverable from the filename or the document itself, record it as unknown rather than inferring it. Note any file the harness cannot read (no extractable text, unsupported format) as such: it counts against coverage, not silently out of the census. Record the total — the census is what makes coverage and staleness checkable later.
4. Choose the sample. All of it up to ~12 pieces. Beyond that: weight toward recency (voices move), then deliberately cover the range — the longest and shortest, each distinct genre or series, and at least one early piece so drift is visible. Record inclusions and exclusions with reasons.
5. Reserve the hold-out. Set aside one piece — two if the corpus has ten or more — as the calibration target: recent, representative, not an outlier. Record only its title, subtitle, date, URL, and one-line topic. Do not read it, do not fetch it, and do not let a subagent fetch it. Under refresh, the newest unread piece is the natural hold-out and doubles as a drift test.
6. Write profile/provenance.md (census, sample, exclusions, hold-out, accessed dates, mode, consent line) and stub profile/index.md.

Phase 1 — Close reading

For each sampled piece, read it in full and write profile/corpus/<piece-slug>.md per references/corpus-note.md: the structural map (how it opens, how sections are divided and headed, how it ends), its thesis in one sentence, the moves it makes, apparatus used, length, and two or three short exemplar excerpts with what each demonstrates. Append a line to profile/corpus/index.md.

Notes are per-piece and evidential. Do not generalize here — the temptation to write "the author always…" after two pieces is what Phase 2 exists to discipline.

Phase 2 — The voice print

Synthesize profile/voice-print.md, profile/structure-repertoire.md, and profile/lexicon.md from the notes, following references/voice-print.md, which defines what to measure and how to record it. In outline:

• Measured features — sentence length (median and range), paragraph length, section count and length, piece length distribution, first-person rate, question rate, punctuation rates per 1,000 words (em-dash, colon, semicolon, parenthesis), apparatus counts. Each with the sample it was computed over, and with the measurement method stated (counted exactly, or estimated from sampled passages — both acceptable, conflating them is not).
• Register and syntax — described, each description anchored to at least one quoted instance: diction level, sentence-shape habits, how emphasis is achieved, how abstraction and concretion alternate.
• Signature moves — with measured rates and one exemplar each; plus the anti-signature list: what the corpus never does (and the ration guidance for each signature, per principle 7).
• Structure repertoire — opening patterns (with which pieces used which), section grammar, ending patterns, title and subtitle grammar, apparatus norms, and typical length by piece type.
• Lexicon — recurring terms and the sense the author gives them (a persona's vocabulary is often a small conceptual toolkit reused across pieces), plus words and constructions conspicuously absent.

Then profile/exemplars.md: at most a dozen excerpts, each ≤ ~40 words, chosen because a drafter can calibrate against them — an opening, a transition, an emphatic close, a concession, a handling of evidence — each with its source and what it demonstrates.

Phase 3 — Positions, continuity, venue

Three files that make a persona more than a style:

• positions.md — the recurring theses and commitments, each stated in the author's terms, quoted, and cited to the piece it comes from. Distinguish held positions from recurring preoccupations (a subject returned to without a settled view). Under any mode but self, this file is quotation-bound: no extrapolated views, no "would probably argue."
• continuity.md — a chronological record of what has been argued where, so a new piece can cross-reference prior work accurately ("as I argued in X"), avoid re-arguing settled ground, and notice when it contradicts the corpus. One entry per piece: date, title, URL, thesis, what it established, and any vocabulary it introduced.
• audience-and-venue.md — the platform's mechanics and the readership's expectations: publication cadence, length norms, sections or series, paywall behavior, citation and footnote practice, how comments and replies are handled, what the venue's tagline claims. Note where the venue's norms and the author's habits diverge, since a drafter must satisfy both.

Phase 4 — Calibration (skip if calibration=false)

The falsification step. Follow references/calibration-protocol.md:

1. From the hold-out's title, subtitle, date, and one-line topic only — plus the profile you just wrote — draft ~400–600 words of its opening as the persona would. Save it as the prediction, before reading anything.
2. Now read the real piece.
3. Compare: the measured features side by side, then the qualitative divergences — the move the profile did not predict, the structure it got wrong, the register it overshot, the tic it over-produced.
4. Correct the profile where the hold-out shows it wrong, and record every correction with its evidence.
5. Where the harness supports subagents, add the blind check: give a fresh context the two openings unlabeled and ask which is the author's and why. The reasons are the signal — record them verbatim as diagnostic; do not treat the verdict as a score.

Write profile/calibration.md: the prediction, the measured comparison, the divergences, the corrections made, and a plain statement of how close the profile got. If the hold-out reveals the profile is badly off, say so in index.md — a profile that knows its own error is usable; one that hides it is not.

Phase 5 — Package and register

1. profile/index.md — the hub every consuming skill loads first: persona name and mode (with usage restrictions if any), corpus size and coverage fraction, bootstrap or refresh date, staleness verdict per the convention, a one-paragraph characterization of the voice, the file list with one line each on when to load it, the calibration verdict, and the known gaps.
2. Append to PERSONA.md a short, marked note: profile built, date, corpus coverage, and anything the human should decide (a mode that was assumed, a tic the corpus shows that they may not intend, a genre the profile cannot cover). Never rewrite the human's seed prose.
3. Report to the operator: corpus size and how much was read, the three most distinctive things the profile found (with a quoted instance each), the anti-signature list's headline items, the calibration result stated honestly, the assumed or confirmed usage mode with its restrictions, the staleness rule that now applies, and the paths. Name the natural next step — draft-from-outline with persona: <persona-id> — and state plainly that the profile constrains and measures voice rather than reproducing it (the limit is in the convention's closing section; do not oversell).