Slack style builder

Produces ~/.claude/slack-voice/voice.md: a distilled profile of how the user actually writes in Slack, plus a bank of their real messages to imitate. slack-cli reads that file before every send/reply.

Why this exists

Agent-written Slack messages are ~100x longer than what the user types by hand, and they read like a report rather than a colleague. Relaying LLM output verbatim pushes work onto the reader and reads as unprofessional (nomeatproxy.com). The honest fix is to read, verify, and answer in your own words. This skill keeps the verification and automates the "own words" part, by learning the words from the user's own history.

Run it

slack style scrape --since 180d        # fetch + label; incremental, safe to re-run
slack style stats                      # evidence report (also written to stats.md)
slack style exemplars -n 200           # best hand-typed messages, with context

scrape takes a few minutes on a first run (paged search + one conversations.replies/history call per thread/channel, with 429 backoff). Re-runs skip everything already in the corpus. --full rebuilds from scratch.

How the labeling works

Exact, not guessed. Slack tags every message posted through the Web API with app_id/bot_id; a message typed in a Slack client has client_msg_id and neither. So provenance: human is genuinely the user's own writing and provenance: app is everything an agent posted as them. search.messages does not return those fields — only conversations.history/replies do — which is why the scrape does two passes.

One blind spot: text pasted from an agent into the Slack client is posted by the client, so it labels as human. These are easy to spot in the exemplar dump — they carry the formatting tells (bullet lists, bold section headers, "Everything is already set up correctly") that the rest of the corpus never uses. Hand-pick past them when writing the exemplar section; do not just take the top N.

Then write the profile

Read slack style exemplars -n 200 output and stats.md, then write ~/.claude/slack-voice/voice.md with these sections:

1. Register — the output register is ASD-STE100 Simplified Technical English, not the register the corpus measures. Sentence case, no contractions, keep every article, one instruction per sentence, simple tenses, active voice, and only the modals can / will / must. Write that rule set into the profile first. Then still record what the real messages do — capitalization, contractions, recurring openers and closers, emoji habits, how they address people, how they say "I'm on it" / "that's done" / "I was wrong" — but record it under a heading that marks it superseded evidence. The corpus remains the source of length, structure and anti-patterns. It is no longer the source of wording. Full rule set: slack-cli/SKILL.md → "Register: ASD-STE100 Simplified Technical English".
2. Length — the measured median and p90 from stats.md, stated as the target.
3. Escape hatch — what to do when detail is genuinely needed: verdict in the message, detail in a thread reply or behind a link, and say what was left out. State explicitly that the link must be something the recipient can open — a PR, a dashboard, a shared doc, a Slack file upload — and that a path on the sender's own machine (~/…, /tmp/…) is never an acceptable pointer. An agent drafting a message knows its own filesystem and will reach for it otherwise.
4. Anti-patterns — drawn from the app bucket, which is a ready-made list of tells: em dashes, *Bold lead-in.* openers, bullet walls, stacked "two things" structure, unprompted jargon.
5. Exemplars — 15–25 real messages grouped by intent (ack, on-it, status, unblock-confirm, ask, correction, defer, escalate), each with the message it was answering. Give every one a paired STE rewrite, and label the original superseded — copy the length and the structure, not the wording. The rewrites are what the agent imitates. This section does the real work; rules alone do not reproduce a register, but a bank of worked rewrites does.

Re-run the whole thing when the profile drifts or after a few months of new history.

Privacy — read before touching git

The corpus is real conversations with real colleagues, and this repo is public.

• Everything the builder writes lives in ~/.claude/slack-voice/ (corpus.jsonl, state.json, stats.md, voice.md). Never copy any of it into this repo, a commit message, a PR body, or an issue.
• Ship only code and instructions here. If an example is needed, invent one.
• SLACK_VOICE_DIR overrides the location (used by the tests).