ohad6k/ditto

O que ditto faz?

Turn the user's own AI session logs into a you.md they can drop into .claude/skills/, .codex/skills/, AGENTS.md, or cursor rules. Runs locally, redacts secrets by default. This skill orchestrates the whole flow so the user never touches python or a manual step.

The one rule that makes this ditto (read first)

The profile MUST be mined from the user's raw session logs (the .jsonl history). NEVER build it from their AGENTS.md, CLAUDE.md, cursor rules, memory files, or a description they type — that just reflects their existing rules back at them, which is the exact opposite of the point and something /init already does. The whole value of ditto is pulling out what they never wrote down, straight from how they actually worked.

So: run ditto.py on the real logs first. If it finds no logs, STOP and ask where the logs are (--path <folder>). Do not fall back to summarizing instructions. No mined logs → no profile. If you ever produce a profile without having run the extractor over real session .jsonl, you have failed the task — say so instead of shipping a fake.

Steps

1. Extract (deterministic, safe). Run the extractor. It reads the user's session logs, keeps only their own words, and redacts API keys / tokens / emails / phone numbers BEFORE anything is written. Never skip redaction.

   First locate ditto.py. Check in this order:

   1. <skill-dir>/../../ditto.py (skill running from inside the cloned repo, at skills/ditto/)
   2. ./ditto.py in the current working directory
   3. Not found → download it: curl -O https://raw.githubusercontent.com/ohad6k/ditto/main/ditto.py (or clone the repo). It is a single stdlib-only file with zero network calls at runtime. (On Windows PowerShell use curl.exe -O; ~ means %USERPROFILE%.)

   Check the runtime: try python --version, then python3 --version, and use whichever works. If neither exists, STOP and tell the user to install Python 3. Do NOT improvise by reading the .jsonl logs yourself — that skips redaction and their secrets end up in the corpus.

   Extraction auto-collapses verbatim-duplicate long messages (re-pasted specs, injected AGENTS.md/CLAUDE.md rules) — this is usually ~40% of a heavy corpus and cutting it saves the user real tokens with no signal lost. Short repeats (approvals like "yes") are kept. The dry-run reports the deduped token count; size chunks off that.

   First run python ditto.py --dry-run and read the token count. Size the chunks so each is ~70K tokens — an agent can genuinely read all of that; 150K-token chunks get skimmed and the profile comes out shallow, while shredding a small corpus into tiny chunks produces noise. Compute chunks = round(tokens / 70000), clamped to at least 4, then run:

   python ditto.py --chunks <N> --out ditto-out
   

   If the corpus is under ~300K tokens, tell the user up front the consensus will be thinner (fewer independent reports) but still real. Depth beats token efficiency on a first mine. Do not reduce the chunk count to save tokens. If no logs are found, STOP and ask where their logs live (pass --path <folder>). Do not proceed without real logs, and never build a profile from their instructions instead. The counts it prints (sessions / messages / tokens / redactions) are your proof it actually read the logs — report them; if sessions is 0, you have nothing to mine.

2. Mine (fan-out). Spawn one sub-agent per file in ditto-out/chunks/, each running the per-chunk prompt from MINING_PROMPT.md on its chunk (if not next to ditto.py, fetch it from https://raw.githubusercontent.com/ohad6k/ditto/main/MINING_PROMPT.md). Run them in parallel. Each returns a structured profile of one slice.

   • Keep it cheap. Per-chunk mining is extraction grunt-work — use a fast/cheap model for the sub-agents. Save the strong model for the reduce (step 3), where the judgment actually happens. Don't fan out expensive models across every chunk.
   • If the environment can't fan out sub-agents, process chunks sequentially instead — same prompt, one at a time.

3. Reduce. Merge all the chunk reports with the reducer prompt in MINING_PROMPT.md. Rank every trait by how many chunks independently surfaced it — high-frequency traits are the real person, one-offs are noise. Every kept trait carries its count inline ((18/20)) — that's the receipt. Output a lean you.md (deep quotes go in a separate you-appendix.md, not in you.md), plus ditto-out/card.json per the reducer spec.

4. Install + prove. The you.md already starts with name: / description: frontmatter, so it's skill-ready. Place it where the user's agent actually reads it, per their tool, then verify it registered:

   • Claude Code → save as .claude/skills/you/SKILL.md. Confirm it shows in the skill list (invoke /you). For it to load in every project, put it in the user-level ~/.claude/skills/you/SKILL.md.
   • Codex native skill - save as ~/.codex/skills/you/SKILL.md. Confirm it appears in Codex's available skills, or open a new Codex session and verify you is loaded before relying on it.
   • Codex repo context - append the body (everything below the frontmatter) to AGENTS.md at the repo root. Codex reads it automatically, no frontmatter needed. Use this when the profile should apply only to one repo.
   • Cursor → save as .cursor/rules/you.mdc, with frontmatter description: act like me and alwaysApply: true, then the body.
   • Gemini CLI → append the body to GEMINI.md. Then prove it with a fixed probe: with the profile loaded, draft a 3-line reply or commit message in the user's voice and ask them "does this sound like you?" Do not claim it's installed until you've confirmed the agent actually picks it up.

5. Show the card. Run python ditto.py --card — it renders their profile card in the terminal and writes ditto-out/card.html (a screenshot-ready card: archetype, top laws with receipt counts, session stats, the one uncomfortable truth). Tell the user this is the shareable piece — the card, never the full you.md.

Rules

• Local only. No network calls. The user's logs never leave their machine. Say so.
• Redaction is not optional. If the user asks for --no-redact, warn them their secrets will appear in the corpus and any shared output.
• Don't invent traits. Every line in you.md must trace to something they actually wrote in the logs. Cut generic filler.
• Quality bar before you call it done: every law in you.md must have at least 2 verbatim quotes behind it in you-appendix.md, and no trait may be something a stranger could have guessed ("wants working code" = filler, cut it). If the profile reads generic, the mine failed — re-run the weak chunks, don't ship it.
• Mine the logs or stop. The profile comes ONLY from the raw session .jsonl. Never synthesize it from AGENTS.md / CLAUDE.md / memory / a description — that is not ditto, it is their rules file reflected back.
• Keep you.md lean — it loads on every task.