Human Distill

Human Distillation

Turn a person's source material into a working coworker agent: a persona written into the agent identity files (IDENTITY.md, SOUL.md, USER.md, CV.md) and a work-module skill (skills/<alias>-playbook/) that carries their workflows, preferences, and judgment. Every generated claim cites the corpus documents it came from; nothing is invented beyond the source.

The deterministic engine is the hybridclaw coworker CLI. Your judgment enters the pipeline through exactly one artefact: extraction.json. You never edit the persona files directly — the engine renders them from validated claims so every line stays cited, versioned, and reversible.

Hard rules

1. Consent first. Distilling a real, named human requires a recorded consent artefact. Never run coworker consent record on your own initiative or invent a consent statement — the operator must provide the statement and run (or explicitly dictate) the command. If a run is blocked, relay the remediation message and stop.
2. Evidence or nothing. Every claim in extraction.json must cite real corpus document ids from the analysis packet. If you cannot support a claim, leave it out or put the question in openQuestions. The engine flags and drops uncited claims — do not try to route around it.
3. Never impersonate. The coworker mirrors the subject's judgment, not their identity. Do not sign as the subject or present generated output as written by them.
4. Privacy. Third-party PII is masked at ingest. If you see unmasked third-party contact details anywhere in generated output, stop and run hybridclaw coworker eval --alias <alias> to surface it.

Pipeline

hybridclaw coworker distill --alias <alias> --name "<display name>" \
  [--role "<role>"] [--match-alias <name|email>]... --source <path> [...]

Stages run in order, each resumable: ingest → analyse → build → merge → correct. The run record lives at runtime/distill/<run-id>/run.json in the coworker's agent workspace, with a human-readable REPORT.md beside it.

1. Intake. Ask the operator: who is being distilled (display name, role, relationship), which aliases/emails identify their authorship (--match-alias, critical for chat exports), whether they are a real person (default) or fictional (--fictional), and what source material exists.
2. Consent. For a real person, confirm consent is recorded (hybridclaw coworker consent show --alias <alias>). If not, give the operator the exact consent record command to run and wait.
3. Ingest + analyse. Run coworker distill with the sources. The engine masks third-party PII, weights quality (authored long-form > chat one-liners), holds out a slice for eval, and writes an analysis packet.
4. Extract. Read analysis/PACKET.md in the run directory and write analysis/extraction.json following references/extraction-contract.md and the six-dimension model in references/six-dimensions.md.
5. Merge. Resume the run (coworker distill --alias <alias> --resume <run-id>). The engine validates citations, merges claims, writes the persona files and the work-module skill as versioned edits, and opens review items for any conflict you declared. Report flagged claims and open reviews to the operator verbatim from REPORT.md.
6. Verify. Run hybridclaw coworker eval --alias <alias>. A leakage failure must be reported and fixed before the coworker is used.

Intake modalities

Use every channel of evidence the operator can provide; they compound:

After each merge, check openQuestions and dimension coverage in the packet; offer the operator a fresh interview round for the weakest dimensions.

Incremental updates and corrections

• New material later: same coworker distill command — only the delta is re-analysed; standing conclusions are never overwritten. Contradictions you declare via conflictsWith become review items the operator resolves with coworker review resolve.
• When the operator corrects the coworker's behaviour in conversation ("she'd never open with a greeting"), persist it immediately: hybridclaw coworker correct --alias <alias> --note "<correction>". It becomes a maximum-weight corpus document and is promoted into the persona on the next run. Then continue with the corrected behaviour in the current session.

Operating boundaries

• Green: reading sources/corpus/status/reports, generating questionnaires, writing extraction.json, coworker status|eval|interview|review list.
• Amber (confirm with the operator first): coworker distill runs and resumes (they write workspace files, reversibly), coworker correct, coworker review resolve, coworker export.
• Red (never): recording or fabricating consent, coworker forget (operator-only), editing generated persona/skill files by hand, distilling someone the operator has not named.

Container note

If hybridclaw is not on PATH (sandboxed container session), do the file-contract half yourself — read PACKET.md, write extraction.json — and hand the operator the exact distill --resume command to run locally.

Deeper material

• references/six-dimensions.md — the persona model and what evidence each dimension needs
• references/extraction-contract.md — the extraction.json schema with a worked example
• references/interview-protocol.md — running subject and colleague interviews well
• references/mirroring.md — the live mirroring loop and fidelity grading