garden-voice

Builds and maintains ~/garden/meta/voice.md: the user's voice profile derived from their own messages. Loaded on-demand by drafting tasks for tone-matching.

Why this exists

Asking a user to describe their writing voice in the abstract gives shallow results. Sampling their actual messages and synthesizing patterns gives the agent a concrete, calibrated reference. Per the source video, having a voice profile is reportedly the single biggest output-quality multiplier for any drafting work.

Modes

init

First-time bootstrap. Sample broadly across the user's recent sent messages.

1. Find user ID via slack_search_users (use the user's name or email from meta/user.md).
2. Pull sent messages. Use slack_search_public_and_private with query from:<USER_ID>. Paginate to gather ~500–1000 messages. Skip threads where the only content is reactions or single-emoji replies.
3. Aggregate text into a single corpus.
4. Redact (see redaction rules below): strip proper nouns, numbers, URLs, but preserve all linguistic content.
5. Synthesize patterns: observe and write down:

• Sentence rhythm (terse / flowing / fragments)
• Capitalization habits (lowercase casual / sentence case / mixed)
• Punctuation tics (em-dashes, ellipses, parenthetical asides, trailing ? for soft asks)
• Vocabulary preferences (favored words, slang, technical register)
• Emoji and reaction patterns (which, how often, in what positions)
• Greeting / closing patterns
• Sentence starters and connector words
• Tone markers (irony, hedging, directness)

1. Pick 10–15 anchored examples: short snippets (≤2 sentences each, redacted) that demonstrate the patterns concretely.
2. Write ~/garden/meta/voice.md per the output format below.
3. Commit to the vault git repo with message voice: init from <N> Slack messages.

refresh

Periodic update: appends new observations rather than overwriting.

1. Read existing voice.md to know what patterns are already noted.
2. Pull last ~200 messages since the file's updated: date.
3. Diff: any new patterns or notable shifts in style?
4. Update voice.md: refresh anchored examples (drop oldest, add freshest), append new pattern observations.
5. Commit with voice: refresh: <summary of changes>.

add <snippet>

Manual addition. The user gives a representative message that should be preserved.

1. Append to the "Anchored examples" section of voice.md with light redaction.
2. If it surfaces a pattern not yet noted, add it.
3. Commit with voice: add: <one-line summary>.

Output format: meta/voice.md

---
type: meta
updated: YYYY-MM-DD
sample_size: <N messages>
sample_window: <date range or "all-time">
---

# Voice profile

Derived from <N> Slack messages sent by the user.

## Patterns

### Sentence structure
- <observation 1>
- <observation 2>

### Capitalization & punctuation
- <observation>

### Vocabulary
- Favored words: <list>
- Avoided words: <list>
- Register: <casual / formal / mixed by audience>

### Tone markers
- <observation>

### Emoji & reactions
- <observation>

### Greetings / closings
- <observation>

## Anchored examples

Short, redacted snippets demonstrating the patterns. `<NAME>`, `<COMPANY>`, `<NUM>`, `<URL>` are placeholders.

1. "<example 1>"
2. "<example 2>"
...

Redaction rules

The voice profile captures style, not content. Apply these substitutions before writing examples:

Preserve unchanged: vocabulary, punctuation, sentence structure, emoji, line breaks, slang, profanity.

If a message can't be meaningfully redacted (i.e. removing the proper nouns leaves nothing distinctive), skip it.

Privacy & safety

• voice.md lives in the private vault. Never copies of it land in the public gardenkit repo.
• Only the user's own sent messages are sampled: never quote others.
• Don't pull DMs unless the user explicitly opts in for a future variant.
• If a Slack message contains obvious secrets (credentials, tokens), drop the message entirely.

How voice.md gets used

meta/soul.md instructs the agent: "When drafting in the user's voice (Slack, email, PRs, tweets, blog posts), reference `[[meta/voice]]` first."

Voice is load-on-demand, not always-on. The SessionStart hook doesn't inject it (would bloat context). The agent loads it only when a drafting task surfaces.

Don't

• Don't sample messages from before the user joined a workspace they care about.
• Don't include verbatim examples longer than 2 sentences.
• Don't synthesize more than 15 anchored examples: beyond that, agent context-load gets unwieldy.
• Don't push voice.md to a public repo (sanity-check git remote -v shows the private vault before committing).