personal-voice-capture

Personal Voice Capture

Teach an AI agent to write like you — not from a one-off prompt, but from your own writing and your own edits.

Every agent I use writes the same way. Competent, fluent, and unmistakably not me. I bet yours do too.

Personal Voice Capture is a Claude Code meta-skill that compiles your real writing into a durable, reusable voice skill. Run one short calibration session: the agent reads your samples, tests what it learned against your edits, and writes a SKILL.md that any future agent loads to draft or rewrite in your voice — chat, email, memos, public posts. You stop re-explaining "make it sound more like me" every time.

It captures how you express things, not what you believe. A voice skill rewrites the intent you give it. It doesn't decide what you think.

Why

Generic AI prose is recognizable, and it's flat. The usual fix — paste a few examples, hope for the best — doesn't compound. The next session starts from zero.

So treat your voice as something to compile from evidence, not guess from a prompt:

your writing samples  +  generated probes  +  your edits
        →  voice rules + approved examples  →  a reusable SKILL.md

The output is portable. Build it once, then load it from any agent, in any project.

How it works

The meta-skill runs an interactive calibration loop:

1. Scope & consent — whose voice, which languages, which contexts (DM, email, memo, post…), and where the output skill lives.
2. Ingest — you hand it writing samples: pasted, files, or exports you authorize.
3. Draft a voice model — a first hypothesis covering cadence, diction, structure, warmth, and context-specific habits.
4. Calibrate — the agent shows you generic-vs-personalized rewrites in an editable file. You edit them inline.
5. Diff into rules — your edits become small, testable voice rules: add, strengthen, weaken, delete, or split by context.
6. Consolidate — every few rounds the model gets rewritten and re-tested.
7. Emit a voice skill — once the voice stabilizes, it generates a <your-name>-voice skill (e.g. eran-voice) you keep using.

It's not a style preset. It's your voice, compiled from your own edits.

See the difference

Same intent, two voices. The first is what an agent writes by default. The second is the same paragraph after a voice skill is loaded (here, the author's own).

Generic:

In today's fast-paced environment, it's more important than ever to ship quickly. Building a small prototype over a weekend can help validate an idea before investing significant resources. We encourage everyone to experiment and learn by doing.

In voice:

People keep waiting for permission to build.

You can just build things now.

I shipped a working prototype over a weekend, by myself. Two days gave me the answer a quarter of planning kept circling: whether the idea was real.

Build the small thing. Let it argue back.

The meaning didn't move — only the expression did.

Install

As a Claude Code plugin (recommended)

/plugin marketplace add eranshir/personal-voice-capture
/plugin install personal-voice-capture@personal-voice-capture

Then invoke it in any session with /personal-voice-capture:personal-voice-capture, or just tell Claude to "capture my voice."

As a standalone skill

Copy the skill directory into your skills folder:

git clone https://github.com/eranshir/personal-voice-capture.git
cp -r personal-voice-capture/skills/personal-voice-capture ~/.claude/skills/

It's then available as /personal-voice-capture.

Usage

Start a session and say something like:

• "Capture my voice."
• "Learn how I write and build my voice skill."
• "Make future agents write like me."

Want to sharpen an existing voice skill later? Correct a draft:

• "That doesn't sound like me — update the voice skill."

The agent runs a calibration round, diffs your edits into rules, and patches the skill.

What you get

A real SKILL.md named <person-slug>-voice that holds:

• a high-level voice signature and per-context modes,
• diction and cadence rules with a phrasebank,
• your own approved before/after examples,
• a transformation workflow and quality checklist,
• a hard rule that preserves your supplied intent and never fabricates your beliefs.

Privacy

Privacy is the default here, not a setting you go find:

• Raw source text isn't retained unless you ask for it. The skill keeps derived rules and the examples you approve — nothing else.
• Calibration files are written to a gitignored scratch path (.scratch/), so source material never lands in version control by accident.
• The generated voice skill encodes style, not private content.

Repository layout

.claude-plugin/plugin.json               # Plugin manifest
.claude-plugin/marketplace.json          # Marketplace catalog (install via /plugin marketplace add)
skills/personal-voice-capture/SKILL.md   # The meta-skill
templates/generated-voice-skill.md       # Template for the per-person output skill
examples/calibration-round.md            # Example calibration interaction
docs/design-notes.md                     # Design rationale and storage policy
scripts/validate_skill.py                # Lightweight SKILL.md frontmatter validator

Development

Validate the skill frontmatter:

python3 scripts/validate_skill.py

If you have the Claude Code CLI, validate the plugin and marketplace manifests too:

claude plugin validate .

Contributing

Issues and pull requests welcome. Good first contributions: more calibration example templates, new context modes, and language-specific voice-capture guidance.

License

MIT © Eran Shir