learn-writing-style

name	learn-writing-style
description	Scan sent messages across configured Google and Slack accounts to derive a personalized writing-style profile (email vs Slack, by audience size), then update GEMINI.md §6 in place. Run after initial bootstrap or anytime you want to refresh.

/learn-writing-style

Derive your writing-style profile from your real outbound messages and write it into GEMINI.md §6. The drafter skills (/draft-follow-up, /daily-brief, /process-inbox) read §6 as the source of truth, so updating §6 propagates automatically.

This skill is interactive: it shows the derived profile and waits for confirmation before writing.

Inputs

• $1 (optional): sample size per channel (default 100; clamped 30–500). Larger samples yield more stable patterns at the cost of longer runtime.
• $2 (optional): --dry-run — compute and show the profile but skip the file write.

Procedure

Date check first. Run shell date "+%Y-%m-%d" to anchor "recent" before any other work.

1. Inventory channels

From GEMINI.md §11, list configured google_* accounts and slack_* workspaces. Skip any whose MCP server is unreachable; note skipped channels in the final report.

2. Sample sent email — per Google account, in parallel

For each google_* MCP, fan out:

• gmail_search_emails with query in:sent newer_than:90d, capped at $1 results.
• For each match, gmail_read_email to retrieve the full body.

Strip quoted reply-history (lines beginning with > or below On <date>, <person> wrote:) before analyzing — only the user's own freshly-typed text counts.

3. Sample sent Slack — per workspace, in parallel

For each slack_* MCP, fan out:

• slack_conversations_search_messages with query from:<self> after:<today minus 90 days>, capped at $1.
• For each match, capture: channel id, channel type (im / mpim / private_channel / public_channel), thread-position (top-level vs reply), and message body.

4. Classify Slack samples by audience size

• Small audience: DMs (im), group DMs (mpim), and any thread reply in any channel.
• Large audience: top-level posts in public channels with num_members > 20, OR any channel name matching common broadcast patterns (announce, general, all-, company-).
• Medium: everything else. If the small or large bucket has fewer than 10 samples, fold the medium bucket into the smaller bucket to stabilize the analysis.

5. Analyze each bucket

Compute these markers per bucket (email, slack-small, slack-large):

• Length: median and p90 word count.
• Capitalization: % of messages whose first character is an uppercase letter.
• Sentence form: % of messages ending in ., ?, or !.
• Sign-off (email only): parse the last 1–2 non-empty lines; cluster by closer pattern (Best, <name>, Thanks, <name>, Cheers, <name>, first-name-only, no-closer). Report top 3 with frequency.
• Greeting (email only): parse the first non-empty line above the body; cluster as in sign-off (Hi <name>,, Hey <name>,, <name>,, no-greeting).
• Em-dash usage: count — per 100 messages. If 0, treat as a hard "never use" rule.
• Markdown usage (Slack only): count messages using *bold*, _italic_, backtick code, bullet/list markers. Report % of messages using each.
• Paragraph format (email only): % of messages where every paragraph is one unbroken line vs % with hard wraps mid-paragraph.

6. Compose the profile draft

Render results in GEMINI.md §6 shape. The exact bullet structure must match what's already there so re-runs are clean overwrites:

## 6. Writing assistant guidance

- **Drafting voice (general):** {{summarized: e.g. "direct and terse, ~M words median per message" or "complete sentences, friendly register"}}.
- **Em-dashes:** {{if count ≈ 0: "Never use em-dashes (`—`) in any drafted communication. Substitute a colon, hyphen, parentheses, or split the sentence."}} {{else: "Em-dashes appear ~K per 100 messages; use sparingly and only when natural."}}
- **Email style:** {{capitalization observation}}; {{sentence form observation}}; sign off with `{{top closer}}` ({{top-closer frequency}}; alternates: {{closer #2}}, {{closer #3}}); {{greeting observation}}; {{paragraph format observation, including the Gmail hard-break caveat if single-line is dominant}}.
- **Slack style — small audiences (DMs, group DMs, thread replies):** {{tone summary}}; {{capitalization}}; {{markdown habits}}; {{length}}.
- **Slack style — large audiences (broadcast channels, announcements):** {{tone summary}}; {{markdown habits}}; {{length}}.
- **Expanding stubs:** if a note is a one-liner or has `#stub`, offer to expand it using linked context and related notes.
- **Surfacing related notes:** search across `+ Atlas/` and `+ Spaces/` to find genuinely relevant prior thinking before writing anything new.
- **Never invent facts** or create fake citations. If you need a source, say so.

_Generated by `/learn-writing-style` on YYYY-MM-DD from N email + M Slack samples. Re-run to refresh._

7. Show the profile inline and confirm

Print the proposed §6 body in chat, plus per-channel sample counts. Then ask the user to confirm:

• Apply — overwrite §6 with the new profile.
• Let me adjust — accept user edits, then apply.
• Cancel — discard.

8. Write on confirmation

Replace the body of ## 6. Writing assistant guidance in GEMINI.md (overwrite from the heading to the next H2 or EOF). Preserve the section header and everything else in the file. Do not append.

If $2 is --dry-run, skip step 8 entirely.

9. Update memory (if applicable)

If your environment maintains a writing-style memory entry, refresh it to mirror the new §6 profile, citing GEMINI.md §6 as the source of truth.

10. Report

Print:

• Path to GEMINI.md
• Sample counts per channel and per bucket
• Channels skipped due to insufficient data (< 5 samples)
• Footer freshness line

Notes

• Insufficient data per channel: if a channel returns fewer than 5 valid samples, skip it. Note in the profile that this channel was not learned (use the template default for that bullet).
• PII: the skill reads only the user's own sent messages and computes aggregate statistics. Sample bodies are not stored in the vault — only the derived profile.
• Idempotent: re-running replaces §6 in place. The footer line tracks freshness.
• Bootstrap recommendation: new users should run /learn-writing-style once after bootstrap/setup.sh completes and at least one Google account + one Slack workspace are configured. Until then, §6 carries the template defaults shipped with the starter.
• Scope: only generates the four user-style bullets (general voice, em-dashes, email, Slack-small, Slack-large). The "expanding stubs", "surfacing related notes", and "never invent facts" bullets are template defaults; the skill leaves them alone.
• Matching the §6 footer convention: every run leaves a _Generated by /learn-writing-style on YYYY-MM-DD from N email + M Slack samples._ line at the bottom of §6 so future readers know the section is auto-derived and how fresh it is.