nickwinder/synthteam/distill-slack-persona

Slack Distillation

Turn a colleague's Slack history into a distilled persona doc — a natural-language description of what they know, what they believe, and how they decide. The persona docs produced here are the sole input to the ask-colleague and ask-team skills.

A persona doc captures substance, not voice. No verbatim message text, no tone/style mimicry — patterns and positions only.

Shared data directory

Everything lives under ~/.synthteam/ (override with the SYNTHTEAM_HOME env var). This directory is intentionally outside any skill folder so personas and raw dumps survive plugin reinstalls and stay reachable by whichever ask-* skills are installed:

~/.synthteam/
├── assets/<slug>/          # raw Slack dumps — local-only, never commit
│   ├── raw-messages.jsonl
│   └── metadata.json
└── personas/<slug>.md      # the distilled persona doc — the deliverable

<slug> is the colleague's lowercase first name (e.g. alex).

The workflow has three steps

Step 1 is mechanical (a script). Step 2 is the agent-driven distillation. Step 3 is human review. Do them in order — distillation needs the dump, review needs the doc.

Step 1 — Dump their Slack messages

The dump script lives in this skill at scripts/dump-user-messages.js. It needs SLACK_USER_TOKEN in a .env file (see the repo .env.example; loadEnv walks up from the cwd to find it).

First install its runtime deps once:

cd <this-skill>/  &&  npm install

Then run the dump:

node scripts/dump-user-messages.js <slug> [--months=12]

The script resolves <slug> to a Slack user, runs search.messages with from:@<username> over the time window, expands every thread the user touched, and writes raw-messages.jsonl + metadata.json into ~/.synthteam/assets/<slug>/. DMs and multi-person DMs are excluded by design — personas are grounded in public/channel conversation only. It only reads what the authenticating token can already see.

Step 2 — Distill the persona

references/distillation-facets.md is the full operational spec for this step — the multi-agent pipeline, the five facet definitions, the persona doc format, and the quality checks. Read it in full before starting; it is the authority for everything below.

In short: the current session is the orchestrator. It splits raw-messages.jsonl into chunks, fans out worker subagents to extract per-facet findings, runs one reducer subagent per facet to synthesize sections, optionally runs a critic pass, then assembles ~/.synthteam/personas/<slug>.md with frontmatter and the fixed five-facet structure. The pipeline follows the ed3d-basic-agents:doing-a-simple-two-stage-fanout pattern.

Do not improvise the persona from general knowledge of the company or the person — every claim in the doc must be defensible from the dump.

Step 3 — Review the diff

Read the finished ~/.synthteam/personas/<slug>.md before relying on it. Spot-check a few claims against raw-messages.jsonl. If a section feels uncanny — putting words in their mouth that aren't grounded — re-run that facet or trim it. Run the verbatim-leak sweep described in the facets spec: any distinctive 4–6-word substring shared between the persona doc and the raw JSONL is a leak and must be rewritten in indirect prose.

Refreshing an existing persona

Same three steps. The dump script overwrites raw-messages.jsonl (the time window slides forward). The distillation rewrites the persona doc from scratch. Diff against the previous version to see what changed. Monthly is a reasonable cadence.

Privacy

• Raw Slack data stays local. ~/.synthteam/assets/ should never be committed to a repo. Verbatim message text never belongs in the persona doc.
• Persona docs describe what someone believes and how they decide. Even paraphrased, that is sensitive. Treat ~/.synthteam/personas/ as private notes about colleagues. Before adding anyone, consider whether they'd be comfortable with the persona existing.
• The dump script cannot exceed the Slack access the user's token already has, and excludes DMs entirely.

Limitations

• Slack search.messages is user-token only and capped at ~10,000 results per query. A highly active person may hit the cap; the script reports if it did.
• Only public channels and conversations the token's user belongs to are searchable.
• Voice / style mimicry is explicitly out of scope. The persona doc captures substance, not phrasing.