agentsclimarketplace

Leopold learn

Skill Jonhvmp/leopold/skills/leopold-learn

The self-improving charter. Mines past Leopold decisions, session corrections, and git history for recurring judgment calls, skeptic-verifies each candidate, and proposes charter amendments. Never edits CHARTER.md itself — you review and apply.From its SKILL.md

Install
npx -y skills add Jonhvmp/leopold --skill leopold-learn

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

  • 8 stars8 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

4.6 KB, as published. Nobody here has run it

/leopold-learn

The charter is written once, at brief time, from what the user remembered about how they decide. Their actual behavior leaks better data: the calls Leopold logged in DECISIONS.md, the corrections they keep typing in sessions, the reverts in git history. This skill mines those sources, verifies every candidate with a skeptic, and proposes amendments — so the charter gets closer to the real user with every run.

Hard boundary: this skill never edits CHARTER.md itself. The charter is the user's identity; the output is a proposal file the user reviews. You may apply accepted amendments only after the user explicitly picks them.

Preflight

  • The Workflow tool must be available (Dynamic workflows enabled in /config). If not, say so and stop — the mining fan-out is what makes this trustworthy; don't fake it with a single-context pass. Dynamic workflows are a Claude Code runtime feature: on Codex CLI the tool does not exist, so say that plainly and stop rather than degrading to one pass.
  • .leopold/CHARTER.md must exist. If not, point to /leopold-brief and stop.
  • Best after at least one Leopold run (some DECISIONS.md content), but session transcripts and git history alone are enough signal to be worth running.

Step 1 — Gather the sources

Build the workflow's args:

  • projectDir: the absolute repo root.
  • charter: full text of .leopold/CHARTER.md.
  • decisionsPaths: .leopold/DECISIONS.md plus every archived .leopold/runs/*/DECISIONS.md (glob for them). Only include files that exist.
  • transcriptDir: this project's Claude Code session directory — ${CLAUDE_HOME:-$HOME/.claude}/projects/<slug>/ where <slug> is the absolute project path with every / replaced by - (e.g. /home/me/app → -home-me-app). Verify the directory exists and contains .jsonl files; pass '' if not. If there are no Claude transcripts for this project — a fresh install, or the runs happened on Codex — pass '', and say so when you report: the transcript miner is dropped (the workflow's mined N from M source(s) line will show two sources, not three) and the pass mines decisions plus git history only.
  • outPath: .leopold/CHARTER-amendments.md (absolute).
  • maxCandidates: 12 unless the user asks for more/less.

Step 2 — Run the mining workflow

Copy the canonical script reference/leopold-learn.workflow.js from this skill's own folder (the directory holding this SKILL.md, wherever the harness loaded it from) to .claude/workflows/leopold-learn.js in the project (create the dir if needed), then launch it with the Workflow tool:

Workflow({ scriptPath: ".claude/workflows/leopold-learn.js", args: { …Step 1… } })

Do not rewrite the script — it encodes the trust structure: three miners over disjoint sources (decision logs / session transcripts / git history, each blind to the others), a cluster pass where a pattern surfacing in more than one independent source is the strongest signal, one skeptic per candidate that defaults to kill, and a distill pass that writes the proposal file only.

Pass args as real JSON, never stringified.

Step 3 — Present the proposal

When the run finishes, read .leopold/CHARTER-amendments.md and show the user the proposed rules (rule + evidence, compact). Two honest outcomes:

  • No survivors — say so plainly: the charter already covers how this project runs. That is a good result, not a failure.
  • Survivors — for each rule, the user decides: fold it into CHARTER.md (append to the section the proposal suggests — Always / Never / Tie-breakers / Preferences), or discard. Use AskUserQuestion (multiSelect) when there are several; apply exactly the accepted ones with Edit, keeping the charter's existing style. Delete the proposal file afterwards if everything was resolved.

Cadence

Suggest running /leopold-learn after each substantial Leopold run finishes, or whenever the user notices they keep correcting the same thing. Each pass compounds: the better the charter, the fewer escalations and wrong-direction decisions the next run makes.

Keep looking

Skills are one crate of 610,677. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.