--- name: leopold-enhance version: 0.1.0 description: "Control plane for the global prompt enhancer: status, on/off, preview a prompt against the gate, and the learn loop that mines your enhancement ledger into proposed prompt-profile rules. Never edits PROMPT-PROFILE.md itself — you review and apply." allowed-tools: - Read - Write - Edit - Bash - Glob - Grep - Workflow - AskUserQuestion triggers: - leopold enhance - prompt enhancer - enhance my prompts - tune the enhancer --- # /leopold-enhance The enhancer is a global `UserPromptSubmit` hook: weak prompts (short, vague, unanchored) get a structured interpretation from Haiku — on the user's own account — injected next to the raw prompt; strong prompts pass through untouched. Everything lives in the enhancer home `$ENH` below (engine, `state.json`, ledger, `PROMPT-PROFILE.md`). This skill is its control plane. Dispatch on the user's argument: - `status` (or no argument) → **Status** - `on` / `off` → **Toggle** - `preview ` → **Preview** - `learn` → **Learn loop** - anything else → **it's a task brief, not a control verb.** The hook has already gated the argument and, when it scored weak, injected the interpretation above (by design: `/skill ` prompts are gated on the ARGUMENT). Don't run the control plane — just do the task the argument describes, honoring the injected interpretation if one is present. ## Paths Nothing here hardcodes a harness home. Resolve both paths first and reuse them in every command below (`$ENH` is the enhancer's own dir under whichever harness home has it; `$LEO` is the Leopold asset home): ```bash ENH="${LEOPOLD_ENHANCE_DIR:-${LEOPOLD_HOME:+$LEOPOLD_HOME/enhance}}" ENH="${ENH:-$([ -d "${CLAUDE_HOME:-$HOME/.claude}/enhance" ] && echo "${CLAUDE_HOME:-$HOME/.claude}/enhance" || echo "${CODEX_HOME:-$HOME/.codex}/enhance")}" LEO="$(leopold home 2>/dev/null || echo "${LEOPOLD_HOME:-$([ -d "${CLAUDE_HOME:-$HOME/.claude}/leopold" ] && echo "${CLAUDE_HOME:-$HOME/.claude}" || echo "${CODEX_HOME:-$HOME/.codex}")/leopold}")" ``` If `$ENH/enhance.py` does not exist, the enhancer isn't installed — point at `leopold menu` (enhance → Install) or re-running the Leopold installer, and stop. ## Status Run `python3 "$ENH/enhance.py" --event status` and read `$ENH/state.json`. Report: enabled/off, model, mode (safe/normal), consecutive failures (if any, mention the self-heal downgrade), and the ledger — total lines plus the last ~5 entries of `$ENH/enhancements.jsonl` (show `prompt_excerpt`, `score`, `injected`, `latency_ms`). If the ledger has 20+ injected entries and the learn loop has never run, suggest `/leopold-enhance learn`. ## Toggle Prefer the extension manager so the menu and the skill stay one control plane: `bash "$LEO/extensions/enhance/manage.sh" toggle on|off`. If that path doesn't exist (driver-only install), fall back to `python3 "$ENH/enhance.py" --event toggle on|off`. Echo the resulting state. Turning on for the first time also runs a capability probe (safe vs normal mode) — relay what it printed. ## Preview Run `python3 "$ENH/enhance.py" --event preview ""` and show the output as-is: the per-signal breakdown, the score against the threshold, and — when it passes — the exact block that would be injected. This is the tuning tool: if the user disagrees with a verdict, point at the threshold env vars (`LEOPOLD_ENHANCE_MIN_SCORE`, `_MAX_WORDS`, `_COOLDOWN_S`) or the `thresholds` object in `$ENH/state.json`. Preview never writes the ledger. ## Learn loop **Hard boundary: this skill never edits `PROMPT-PROFILE.md` itself.** The profile shapes every future interpretation; the output is a proposal file the user reviews. You may apply accepted amendments only after the user explicitly picks them. ### Preflight - The `Workflow` tool must be available (Dynamic workflows enabled in `/config`). If not, say so and stop — the independent-miners-plus-skeptic structure is what makes this trustworthy; don't fake it with a single-context pass. Dynamic workflows are a **Claude Code** runtime feature: on Codex CLI the tool does not exist, so name that as the reason and stop. The enhancer itself (status, toggle, preview) works on both harnesses — only this learn loop is Claude Code only. - `$ENH/enhancements.jsonl` must exist with at least ~10 `"injected":true` lines. Fewer → say the ledger is too thin to mine yet. ### Step 1 — Gather the sources Build the workflow's `args`: - `ledgerPaths`: `$ENH/enhancements.jsonl` plus `enhancements.1.jsonl` if it exists (absolute paths — expand `$ENH`, the workflow takes real paths, not shell). - `transcriptDirs`: the unique `cwd` values of the last ~200 ledger lines, each mapped to `${CLAUDE_HOME:-$HOME/.claude}/projects//` where `` is the absolute path with every `/` replaced by `-` (e.g. `/home/me/app` → `-home-me-app`). Keep only dirs that exist and contain `.jsonl` files; pass `[]` if none — on a Codex-only machine there are none, so the transcript miner is dropped and you say so rather than reporting a full pass. - `profile`: full text of `$ENH/PROMPT-PROFILE.md`. - `outPath`: `$ENH/PROFILE-amendments.md` (absolute). - `maxCandidates`: 10 unless the user asks for more/less. ### Step 2 — Run the mining workflow Copy the canonical script `reference/leopold-enhance-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-enhance-learn.js` in the current project (create the dir if needed), then launch it: ``` Workflow({ scriptPath: ".claude/workflows/leopold-enhance-learn.js", args: { …Step 1… } }) ``` Do not rewrite the script — it encodes the trust structure: two miners with disjoint lenses (transcript reactions vs ledger statistics, so a pattern surfacing in both is real signal), one kill-biased skeptic per candidate, and a distill pass that writes the proposal file only. Pass `args` as real JSON, never stringified. ### Step 3 — Present the proposal Read `$ENH/PROFILE-amendments.md` and show the proposed rules (rule + kind + evidence, compact). Two honest outcomes: - **No survivors** — say so plainly: the enhancer is reading the user's prompts the way they mean them. That is a good result, not a failure. - **Survivors** — use AskUserQuestion (multiSelect). For each accepted `interpretation` rule, append a `- rule` line to `$ENH/PROMPT-PROFILE.md` with Edit. For each accepted `gate` rule, apply the suggested threshold change to the `thresholds` object in `$ENH/state.json`. Delete the proposal file afterwards if everything was resolved. ## Cadence Suggest `/leopold-enhance learn` once the ledger has a few dozen injected entries, and again whenever the user mentions the interpretations feel off. Each pass compounds: the better the profile, the closer every future interpretation reads to what they actually meant.