persona-forge

persona-forge

Personalize your Claude Code agent from your own usage history.

persona-forge reads your local Claude Code transcripts, distils who you are and how you work into an evidence-grounded persona, and generates a personalized "working-with-you" skill that auto-loads on your coding tasks — tuned to keep you in flow (lead with the result, act on reversible work, confirm only the irreversible, end every turn with one clear next move). Then it keeps learning from your newest sessions so the agent gets better for you over time.

Everything runs on your machine. No transcript ever leaves your computer; the only network call is the LLM you're already using as Claude Code.

---

Install

/plugin marketplace add akhatua2/persona-forge      # or the path/URL to this repo
/plugin install persona-forge@persona-forge

Or point Claude Code at a local clone:

/plugin marketplace add /path/to/persona-forge
/plugin install persona-forge@persona-forge

Requires Python 3 (standard library only). If tiktoken is installed, token counts are exact; otherwise an estimate is used.

Use

Command	What it does
/persona-forge:personalize	Full run: analyze your history → build the persona → generate and install working-with-<you> skill.
/persona-forge:persona	Just build and show the persona (no skill installed).
/persona-forge:evolve	Learn from your newest sessions and surgically update the installed skill.

Optional args on any command: --days N (limit lookback) and --name "Your Name".

/persona-forge:personalize
# ...work for a few days...
/persona-forge:evolve          # folds new corrections & habits into your skill

A quiet SessionStart hook reminds you to run /persona-forge:evolve if the skill hasn't learned in over a week. To turn it off, remove the SessionStart block from plugins/persona-forge/hooks/hooks.json.

How it works

1. scripts/extract_corpus.py (deterministic, local) pulls your messages out of ~/.claude/projects/**/*.jsonl, strips tool output and system noise, and writes one corpus file per project plus a stats.json (median message length, interrupt rate, active hours, top projects, slash-command habits, …).
2. The command fans out subagents to profile you from the corpus, then synthesizes one persona — including an engagement model derived from your data: when the agent should act vs. ask, when it must stop and confirm, and what response shape keeps you moving.
3. It drafts the skill, self-critiques it (grounded? generic? any dark pattern? any fabricated number?), and installs it to ~/.claude/skills/working-with-<you>/SKILL.md.
4. /evolve re-reads only what's new since the last run, mines your recent corrections and drift, and applies minimal edits — versioned, backed up, and logged to CHANGELOG.md.

The design philosophy lives in the bundled engagement-design skill: engagement = genuine momentum, derived per-person, never dark patterns.

What gets written

~/.claude/skills/working-<you>/SKILL.md     the installed, auto-loading skill
~/.claude/persona-forge/latest/persona.md   the readable persona
~/.claude/persona-forge/state.json          version + last-learned watermark
~/.claude/persona-forge/history/            backups of every skill version
~/.claude/persona-forge/CHANGELOG.md        what was learned, when, and why

Privacy

Your transcripts are read locally and turned into local files. persona-forge makes no network requests of its own. Review the generated skill before sharing it — it will mention your projects, tools, and habits.

License

MIT — see LICENSE.