mph-voice-refresh

Files (4)

• SKILL.md
• VOICE.md
• mine_feed.py · scripts/
• mine_voice.py · scripts/

SKILL.md (first lines)

# MPH Voice Refresh

Keeps the [VOICE.md](VOICE.md) corpus alive. Without this, the writer keeps reusing the same "Michael impersonations" until they sound canned.

## When to invoke
- User asks to refresh / update / mine the voice corpus
- User says drafts are "sounding the same" or "you keep using that phrase"
- It's been more than ~3 weeks since last refresh (check the date at the bottom of VOICE.md)

## Two sources, two miners

| Source | Script | What it captures |
|---|---|---|
| Chat transcripts (`~/.claude/projects/*/*.jsonl`) | `scripts/mine_voice.py` | Unfiltered, conversational — tics like `..`, `lol`, `or whatever`, `Ugh.`, mid-sentence self-corrections. The way Michael actually types. |
| Substack RSS feed (`mphinance.substack.com/feed`) | `scripts/mine_feed.py` | Polished, published — paired openers, contradiction-filter constructions, punchy closers. The way Michael writes for readers. |

Both write to the **same VOICE.md**, but into separate sections. Don't mix them — chat voice has tics that don't belong in a Substack draft, and published voice has constructions that would sound forced in chat.

## Flow

1. **Read current state.** Read [VOICE.md](VOICE.md) so you know what's already in the corpus and what's already retired.

2. **Run the relevant miner.**
   - For chat voice refresh (most common — run every few weeks):
     ```
     python skills/mph-voice-refresh/scripts/mine_voice.py --days 30
     ```
   - For published voice refresh (run after Michael ships a new Substack post):
     ```
     python skills/mph-voice-refresh/scripts/mine_feed.py --limit 8
     ```
   - Run both back-to-back if it's been a while.

3. **Review with Michael.** Show the candidate lists. Do NOT auto-apply. Ask which additions to keep and which retirements to confirm. Voice is subjective; the miner is a candidate generator, not the authority.

4. **Apply approved changes.** Append confirmed additions to the appropriate section of VOICE.md:
   - Chat candidates → "Idioms & tics" / "Sentence rhythms" / "Long-form examples"
   - Feed openers/closers → "Published voice — Substack patterns"
   - Confirmed Claude repeats → `RETIRED`
   - Update the "Last refreshed" date at the bottom.

5. **No commits.** Leave the file dirty. Michael decides when/if to commit changes to the alpha-skills repo.

## What counts as "voice-worthy"

The miner looks for plain-string user messages (not tool results, not system reminders) that contain at least one informal marker: multi-dot punctuation (`..` / `....`), `lol`, `lmao`, `tbh`, `ngl`, `or whatever`, `kinda`, `sorta`, `I mean`, `I guess`, `wait`, `actually`, `yeah`, `damn`, `bullsh*t`, etc. It also requires length between 10 and 400 chars so we skip both one-word prompts and pasted essays.

## What counts as "stale"

Any 5-word phrase that appears in 3+ separate session transcripts (counted from assistant messages). The miner is intentionally aggressive here — better to surface a false positive Michael can ignore than miss a real tic. The user has explicitly said: "you end up reusing the same 'quotes' or whatever all the time" — this is what we're hunting.

## Hard rules

- **Never auto-apply.** Always present candidates for review.
- **Never delete the RETIRED section.** It's a permanent ban list.
- **Don't mine SSH-session transcripts** (project paths starting with `ssh-`). Those are remote work, often not in Michael's voice.

From mphinance/alpha-skills, MIT.

More from mphinance/alpha-skills