Skills

Leopold ships a family of Claude Code skills, installed to ~/.claude/skills/. They are authored in the gstack-compatible format (frontmatter + markdown body). The core skills are documented below.

flowchart LR
    B["/leopold-brief<br/><small>Phase 1</small>"] --> R["/leopold-run<br/><small>Phase 2 · loop</small>"]
    B --> W["/leopold-workflow<br/><small>Phase 2 · workflow</small>"]
    R --> S["/leopold-status<br/><small>watch</small>"]
    R --> X["/leopold-stop<br/><small>take back</small>"]
    R --> L["/leopold-learn<br/><small>charter learns</small>"]
    W --> L
    T["/leopold-triage<br/><small>backlog → plan</small>"] --> B
    E["/leopold-enhance<br/><small>prompt enhancer</small>"]
    P["/leopold-persona<br/><small>synthetic customers</small>"]

/leopold-brief

Phase 1. A structured debate that captures the mission and your decision-making, then writes the brief (MISSION, CHARTER, GUARDRAILS, PLAN) to .leopold/.

• Use it to start a mission, or to revise an existing brief.
• Output the four artifacts plus an empty DECISIONS.md.

/leopold-run

Phase 2. Activates the run (writes state.json with active:true) and begins the turn loop. The Stop hook carries it forward; the guard hook locks git.

• Preflight aborts if the brief is missing — run /leopold-brief first.
• Behavior adopts spawned-session mode so gstack skills auto-decide.

/leopold-workflow

Phase 2, the workflow way. Compiles the same brief into a dynamic workflow — a JavaScript harness Claude Code's runtime executes in the background — and runs it. The plan lives in code instead of one context window, so a long run doesn't drift into agentic laziness, self-preferential bias, or goal drift; each item gets an independent adversarial review.

• Compiles PLAN.md into dependency-ordered waves and risk-classifies each item (effort + critical + sensitive) with the driver's keyword rules, then launches the canonical script (reference/leopold-run.workflow.js) via the Workflow tool.
• Preflight needs Dynamic workflows enabled (/config) and an existing brief.
• Git stays locked for free — a workflow can't commit; it stages, you commit.
• Resumable and visible in /workflows; saving the compiled script to .claude/workflows/leopold-run.js makes the brief a re-runnable harness.
• When to reach for it vs /leopold-run: large or parallelizable plans → workflow; short or interactive plans → the loop.

/leopold-learn

The self-improving charter. Mines the decision log (DECISIONS.md + archived runs), this project's session transcripts, and git history with three miners over disjoint sources; clusters the recurring signals (a pattern surfacing in more than one independent source is the strongest); puts a kill-biased skeptic on every candidate; and distills the survivors into .leopold/CHARTER-amendments.md.

• Hard boundary: it never edits CHARTER.md itself — the charter is the user's identity. The human reviews the proposal and applies what sounds like them.
• Cadence: after each substantial run, or whenever you notice yourself repeating the same correction. Each pass compounds: a sharper charter means fewer escalations and fewer wrong-direction decisions next run.
• Requires Dynamic workflows enabled.

/leopold-triage

Backlog triage as a workflow, with the quarantine pattern: the agents that read untrusted item content (issue bodies) have no repo access and only emit structured classifications; the agents that touch the repo (fix planners) see only those structured fields. A prompt injection in an issue can distort at most its own item's classification.

• Input: GitHub issues (gh issue list), a file/directory of reports, or pasted content; dedupes against open PRs and PLAN.md.
• Output: a severity-ranked actionable list, duplicate groups, already-tracked items, and (in fix mode) grounded fix plans for quick wins — which can become PLAN.md items for a /leopold-workflow run.
• Continuous: pair with /loop (e.g. every 6h) for a standing queue.

/leopold-enhance

Control plane for the global prompt enhancer — the UserPromptSubmit hook that has Haiku (your own account) interpret weak prompts, charter-aware.

• status — enabled/off, mode, and the tail of the enhancement ledger.
• on / off — the same toggle as leopold menu → enhance.
• preview "text" — the gate's verdict (per-signal score breakdown) and the exact block that would be injected; the threshold-tuning tool.
• learn — mines the ledger + session transcripts for enhanced prompts you corrected right after (plus statistical gate misfires), skeptic-verifies each candidate, and proposes rules into ~/.claude/enhance/PROFILE-amendments.md.
• Hard boundary: learn never edits PROMPT-PROFILE.md itself — the profile shapes every future interpretation; you review and apply.
• learn requires Dynamic workflows enabled.

/leopold-persona

Synthetic-customer testing, conducted: an evidence-grounded persona walks a declared product flow — perceiving the real interface, reacting in character, journaling every turn — and the run ends with a structured report of bugs, confusion, accessibility problems and friction.

• init — scaffolds the .leopold/persona/ namespace (personas / flows / runs) and the flow template.
• build <id> — compiles a persona contract through persona-contract-builder.
• run <flow> [--persona <id>|all] — walks the flow; the headless twin is leopold persona run (same run tree, driver-conducted, bounds enforced in code).
• Ships with two support skills, vendored verbatim and installed on both harnesses: persona-contract-builder (compiles an evidence-grounded persona-contract/1.0 — claim catalog, source ledger, epistemic boundaries, validation gate) and persona-contract-runtime (enacts a contract one bounded turn at a time — sincerity protocol, anti-drift gate, instrumented results). They are the persona semantics; /leopold-persona conducts around them and never redefines their schemas.

/leopold-status

Read-only dashboard: active or not, plan progress, decisions logged, recent events. Never mutates anything.

/leopold-stop

Clean shutdown. Flips the run inactive so the Stop hook allows the session to halt at the next turn boundary. The blunt alternative is touch .leopold/STOP.

Frontmatter shape

Each skill declares name, version, description, allowed-tools, and triggers, so Claude Code can route to it by intent.