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 from the distill
plugin. If hybridclaw coworker prints the general usage instead, the plugin
is not installed: ask the operator to run hybridclaw plugin install distill
rather than installing it yourself. 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
- 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.
- 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.
- 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.
- 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.
- 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.
- 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.
- 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.
- 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.
- 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.
- 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:
| Modality |
How |
| Chat exports |
Slack export dirs/JSON, generic chat JSONL — --kind auto detects them |
| Email |
.mbox archives |
| Meetings ("listening") |
Speaker-labelled transcripts (Name: text lines) |
| Writing samples |
Markdown / text docs, decision records, posts |
| Questionnaire |
hybridclaw coworker interview --alias <alias> [--audience subject|colleague] --out <file> generates a gap-driven interview targeting the dimensions with least evidence; the answered file is ingested with --kind interview (highest weight) |
| Mirroring |
Live draft-compare-diff sessions per references/mirroring.md; differences become corrections |
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
1---2name: human-distill3description: "Distill a real person into a hireable coworker agent from their source material — chat exports, emails, meeting transcripts, documents, interviews, and live mirroring. Use when the operator wants to clone a colleague, build a coworker from someone's writing, onboard a digital twin, or asks to 'distill', 'clone', or 'mirror' a person. Consent-gated for real humans."4user-invocable: true5metadata:6 source: https://skillmd.com/skills/hybridaione/human-distill7 hybridclaw:8 category: productivity9 short_description: "Distill a human into a coworker agent."10 tags:11 - coworker12 - distillation13 - persona14 - cloning15 - onboarding16 related_skills:17 - skill-creator18 - personality19---2021# Human Distillation2223Turn a person's source material into a working coworker agent: a persona24written into the agent identity files (`IDENTITY.md`, `SOUL.md`, `USER.md`,25`CV.md`) and a work-module skill (`skills/<alias>-playbook/`) that carries26their workflows, preferences, and judgment. Every generated claim cites the27corpus documents it came from; nothing is invented beyond the source.2829The deterministic engine is the `hybridclaw coworker` CLI from the `distill`30plugin. If `hybridclaw coworker` prints the general usage instead, the plugin31is not installed: ask the operator to run `hybridclaw plugin install distill`32rather than installing it yourself. Your judgment33enters the pipeline through exactly one artefact: `extraction.json`. You never34edit the persona files directly — the engine renders them from validated35claims so every line stays cited, versioned, and reversible.3637## Hard rules38391. **Consent first.** Distilling a real, named human requires a recorded40 consent artefact. Never run `coworker consent record` on your own41 initiative or invent a consent statement — the operator must provide the42 statement and run (or explicitly dictate) the command. If a run is43 blocked, relay the remediation message and stop.442. **Evidence or nothing.** Every claim in `extraction.json` must cite real45 corpus document ids from the analysis packet. If you cannot support a46 claim, leave it out or put the question in `openQuestions`. The engine47 flags and drops uncited claims — do not try to route around it.483. **Never impersonate.** The coworker mirrors the subject's judgment, not49 their identity. Do not sign as the subject or present generated output as50 written by them.514. **Privacy.** Third-party PII is masked at ingest. If you see unmasked52 third-party contact details anywhere in generated output, stop and run53 `hybridclaw coworker eval --alias <alias>` to surface it.5455## Pipeline5657```58hybridclaw coworker distill --alias <alias> --name "<display name>" \59 [--role "<role>"] [--match-alias <name|email>]... --source <path> [...]60```6162Stages run in order, each resumable: `ingest → analyse → build → merge →63correct`. The run record lives at `runtime/distill/<run-id>/run.json` in the64coworker's agent workspace, with a human-readable `REPORT.md` beside it.65661. **Intake.** Ask the operator: who is being distilled (display name, role,67 relationship), which aliases/emails identify their authorship68 (`--match-alias`, critical for chat exports), whether they are a real69 person (default) or fictional (`--fictional`), and what source material70 exists.712. **Consent.** For a real person, confirm consent is recorded72 (`hybridclaw coworker consent show --alias <alias>`). If not, give the73 operator the exact `consent record` command to run and wait.743. **Ingest + analyse.** Run `coworker distill` with the sources. The engine75 masks third-party PII, weights quality (authored long-form > chat76 one-liners), holds out a slice for eval, and writes an analysis packet.774. **Extract.** Read `analysis/PACKET.md` in the run directory and write78 `analysis/extraction.json` following79 [references/extraction-contract.md](references/extraction-contract.md)80 and the six-dimension model in81 [references/six-dimensions.md](references/six-dimensions.md).825. **Merge.** Resume the run83 (`coworker distill --alias <alias> --resume <run-id>`). The engine84 validates citations, merges claims, writes the persona files and the85 work-module skill as versioned edits, and opens review items for any86 conflict you declared. Report flagged claims and open reviews to the87 operator verbatim from `REPORT.md`.886. **Verify.** Run `hybridclaw coworker eval --alias <alias>`. A leakage89 failure must be reported and fixed before the coworker is used.9091## Intake modalities9293Use every channel of evidence the operator can provide; they compound:9495| Modality | How |96|---|---|97| Chat exports | Slack export dirs/JSON, generic chat JSONL — `--kind auto` detects them |98| Email | `.mbox` archives |99| Meetings ("listening") | Speaker-labelled transcripts (`Name: text` lines) |100| Writing samples | Markdown / text docs, decision records, posts |101| Questionnaire | `hybridclaw coworker interview --alias <alias> [--audience subject\|colleague] --out <file>` generates a gap-driven interview targeting the dimensions with least evidence; the answered file is ingested with `--kind interview` (highest weight) |102| Mirroring | Live draft-compare-diff sessions per [references/mirroring.md](references/mirroring.md); differences become corrections |103104After each merge, check `openQuestions` and dimension coverage in the packet;105offer the operator a fresh interview round for the weakest dimensions.106107## Incremental updates and corrections108109- New material later: same `coworker distill` command — only the delta is110 re-analysed; standing conclusions are never overwritten. Contradictions111 you declare via `conflictsWith` become review items the operator resolves112 with `coworker review resolve`.113- When the operator corrects the coworker's behaviour in conversation114 ("she'd never open with a greeting"), persist it immediately:115 `hybridclaw coworker correct --alias <alias> --note "<correction>"`.116 It becomes a maximum-weight corpus document and is promoted into the117 persona on the next run. Then continue with the corrected behaviour in the118 current session.119120## Operating boundaries121122- Green: reading sources/corpus/status/reports, generating questionnaires,123 writing `extraction.json`, `coworker status|eval|interview|review list`.124- Amber (confirm with the operator first): `coworker distill` runs and125 resumes (they write workspace files, reversibly), `coworker correct`,126 `coworker review resolve`, `coworker export`.127- Red (never): recording or fabricating consent, `coworker forget`128 (operator-only), editing generated persona/skill files by hand, distilling129 someone the operator has not named.130131## Container note132133If `hybridclaw` is not on PATH (sandboxed container session), do the134file-contract half yourself — read `PACKET.md`, write `extraction.json` —135and hand the operator the exact `distill --resume` command to run locally.136137## Deeper material138139- [references/six-dimensions.md](references/six-dimensions.md) — the persona model and what evidence each dimension needs140- [references/extraction-contract.md](references/extraction-contract.md) — the `extraction.json` schema with a worked example141- [references/interview-protocol.md](references/interview-protocol.md) — running subject and colleague interviews well142- [references/mirroring.md](references/mirroring.md) — the live mirroring loop and fidelity grading