Self-Tuning Slack Voice Writer

AI-drafted Slack messages fail in two ways: they sound like AI (corporate filler, manufactured enthusiasm, "I'm excited to share...") and they come out the wrong shape — broadcast structure jammed into a two-line reply, feature bullets where a colleague expected a conversation. Static "write like me" prompts don't fix either for long: they're built on guesswork rather than evidence, they go stale as your voice drifts, and you end up re-explaining yourself every session or rewriting every draft. This Blueprint builds your voice profile from messages you actually sent, then keeps it accurate automatically.

1. Install the template skill and invoke it. It detects it's unconfigured and runs setup: a 10-question interview (name, sign-off, Slack member ID, spelling, emoji style, banned words, swearing tolerance, tune schedule).

2. With your permission, it mines 20–50 messages you actually sent — filtering out bots and any automation posting under your ID — and derives evidence-based voice rules: openers and closers by context, emoji habits, hedging style, formatting, message length.

3. It generates your personalised skill with your own real messages embedded as canonical voice anchors, packaged for one-click install. Interview answers capture intent; real messages capture truth — where they conflict, it trusts the messages and tells you.

4. Final setup step: a weekly scheduled tune. Each week it reviews your last 7 days of hand-typed messages and always reports four things — its analysis, any amendments it wants to make to the skill, the reasoning behind them, and quoted message evidence. Amendments only ship when a pattern appears in 3+ separate messages, and nothing changes until you click install.

Day to day, every draft is mode-detected first (Reply/DM, Feedback, Showcase, Utility) — so you get the right voice in the right shape, not an announcement template on a thank-you reply.

Drafts that pass the "would I actually send this?" test without a rewrite pass — and a profile that stays accurate as your voice changes, with zero maintenance from you. The tune loop is human-in-the-loop by design: Claude proposes with evidence, you approve with one click.

Ships as a .skill package (zipped skill directory). The template carries {{PLACEHOLDER}} fields; setup fills them, and because the personalised copy shares the skill name, installing it replaces the template cleanly. Installed skills are read-only to sessions, so every change — initial setup and weekly tunes — ships as a new .skill package you approve. That's the safety model, not a workaround: nothing edits your voice profile silently. Evidence bar: 3+ hand-typed messages before any rule changes; fewer than 10 hand-typed messages in the window = report only, no proposal. The miner and tuner exclude automation posting under your own user ID (scheduled reports, monitoring systems, fixed-skeleton status posts, messages addressing you in the second person). This came from live testing: 17 of my 20 most recent "from:me" results were a monitoring system posting as me — without the filter, the profile trains on a cron job. The weekly task runs in local time while the Claude app is open; missed runs fire on next launch. Click "Run now" once after setup to pre-approve the Slack tool permissions, or future runs pause on prompts.

Requires Claude Cowork (desktop). Scheduled tasks and skill installs don't exist on claude.ai web — without Cowork, you can still run the tune manually by saying "tune" to the skill weekly.

Requires the Slack connector with search access. The skill only reads messages you sent.

Voice quality is bounded by your hand-typed Slack volume. Light writers get thinner profiles and frequent "insufficient evidence" weeks — that's the evidence bar working, not a failure.

Updates are never automatic. Every change to your profile requires your explicit install — deliberate, because a bad silent update to your own voice is worse than a stale one.

One profile per person. Share the template; never share your personalised copy — it contains your real messages.