Clone Yourself

Build <first-name>-clone: a local skill that answers as the user, in the first person, with their opinions and their phrasing. A good clone feels like the person, not like a policy document. It has three layers:

Everything stays on the user's machine. Never upload samples, rules, memory files or the generated clone to Edge, a repository or any other service. The scripts use the standard library only and make no network calls.

1. Voice

Run the clone-my-voice workflow, with its local-only rules, to create or reuse their my-voice skill. The clone points to it instead of copying it.

2. Judgment

Ask which folders hold their rules, and for a short list of people, clients and companies that appear in them. Then run:

python3 scripts/extract_rules.py [their paths] --aggressive --redact-words "Ana,Northwind,..." --out ./clone-work

--redact-words removes the names they list, wherever they appear (also inside name-api or name_sync). --aggressive also removes every other capitalised word in the middle of a sentence, except well-known tools and languages. Use both by default: single-word names are what pattern matching misses.

It also reads the usual places by default (~/.claude/CLAUDE.md, project CLAUDE.md and AGENTS.md, Cursor rules, ~/.claude/projects/*/memory/*.md, ~/.codex/AGENTS.md); add --no-defaults to read only what they name. It keeps imperative rule lines, redacts emails, URLs and domains, IP addresses, phone numbers, handles, money, long numbers, file paths and multi-word names, drops rules that only make sense with a private fact attached, dedupes, and sorts rules into decide / judge / report / bans / format / persona. It prints counts only.

If the agent runs on a hosted model, ask before reading clone-work/rules.json or using --show, because that sends the redacted rule text to the model. Offer to let the user read it themselves instead.

Turn the strongest rules (highest count) into a short proposed profile: about 10 rules, grouped, each tagged with its source file name. The user confirms or edits; set "confirmed": true on the ones they keep. Redaction is a safety net, not a guarantee: read every kept rule for names or private facts before confirming.

3. Persona

This layer is what makes the clone feel like a person. Build it from the user's own sent writing (the samples clone-my-voice used) and the persona rules, then ask the user to correct it:

• How they talk: register in chat vs in public, length, capitalization, humour, how direct they are.
• Phrases they actually use: only phrases that appear more than once in their own writing. Never invent catchphrases.
• What annoys them: what they correct, ban or complain about repeatedly.
• What they care about: what they push for in decisions.
• How they push back and how they say no.

No names, no quotes that identify another person, no private facts. Save the confirmed version as clone-work/persona.json:

{"talk": ["..."], "phrases": ["..."], "annoys": ["..."], "cares": ["..."],
 "example": "> **User:** ...\n>\n> **Sam:** ..."}

The example is one short synthetic exchange in their real rhythm.

4. Build

python3 scripts/build_clone.py --rules clone-work/rules.json --persona clone-work/persona.json \
  --name Sam --tasks "startup tasks" --pronoun they --out ~/.claude/skills

Use the agent's skills folder (~/.claude/skills/, ~/.codex/skills/, or the configured Cursor location; ask if unclear). It writes sam-clone/SKILL.md, references/rules.md and references/persona.md from templates/, uses only confirmed rules, refuses to overwrite without --force, and runs the checker.

5. Verify

build_clone.py runs scripts/check_clone.py, which validates the front matter and scans every file for emails, URLs, phone numbers, handles, money, paths and names. Fix every hit; it exits non-zero until clean. Allow the user's own first name with --allow Sam.

Then test side by side: three prompts the user picks, each answered with and without the clone. The clone's answers should be in the first person, short, opinionated and in their phrasing. If an answer reads like a consultant memo, tighten persona.md and rebuild.

Finish with:

✓ <first-name>-clone created locally. Your AI can now work like you. Turn on Edge local ranking (EDGE_LOCAL_SKILLS=1) and Edge offers it whenever a task needs your judgment.