How do I install the Discover skill?

Run npx skillmds@latest add emmanuelrtm/discover in your terminal (requires Node.js), paste this page's agent-chat prompt into Claude, Cursor, or any MCP-connected agent, or download the SKILL.md file and copy it into your agent's skills directory.

Is Discover safe to use?

This skill has not completed SkillMD's automated safety review yet. SkillMD never runs a skill's scripts for you; review the SKILL.md before installing.

Which AI agents work with Discover?

This skill is tagged as working with Claude Code, Claude.ai, OpenAI Codex. SKILL.md is an open format, so most agents that read a skills directory can load it too.

Is Discover free to use?

Yes. Installing skills from SkillMD is free, and the skill stays under its author's original license.

Who published Discover?

EmmanuelRTM (@emmanuelrtm) published this skill. Their other Agent Skills are listed on their SkillMD profile.

Discover by emmanuelrtm — Agent Skill

Name: Discover
Author: EmmanuelRTM

Discover

Run (or re-run) Stage 2 only — Discovery: Ideator plus Parallel Explore-Exploit branches (Solve → Evaluate → Audit → Rank+Select → distilled feedback), for the configured number of iterations. Requires brief.md from /investigate.

EmmanuelRTM cd56dcb 2.6 KB Updated Sep 17, 2026

File contents

/discover — Stage 2: Ideator + Parallel Explore-Exploit

Arguments: $ARGUMENTS

Current state: !python3 .claude/scripts/ledger.py status

Preconditions: resolve the run (--run or ACTIVE_RUN); brief.md and a valid bibliography.jsonl must exist — otherwise stop and point to /investigate. Read B and I from run-config.json (--iterations here overrides I; create missing iterations/iN/branches/bK dirs as needed).

For each iteration iN:

Launch ideator (run dir, iN, B, prior distilled-feedback.md files if any — also from previous /discover invocations). Expect ideas.md + proposals/p1..pB.md. From i2 on the ideator applies top-K retention (p1..pK refine the previous iteration's top-K surviving branches).
Launch ALL B solver subagents IN ONE MESSAGE — one Agent call per branch, each given only its own iterations/iN/branches/bK/ dir and proposals/pK.md. This parallel fan-out is the point of PEE; sequential launches are a protocol violation.
When all solvers are done: launch B evaluator subagents in one message.
Then B auditor subagents in one message.
Rank + Select (you): read all eval.json + audit.md; audit-FAIL branches are disqualified regardless of score; write iterations/iN/ranking.md (table + rationale).
Distilled feedback (you): invoke the distill-feedback skill (Skill tool, skill: distill-feedback) and write iterations/iN/distilled-feedback.md with exactly its section template — never improvise the format.
Ledger: python3 .claude/scripts/ledger.py append '{"event":"stage_iteration_iN","detail":"best=<branch>:<score>"}'

After the final iteration: select the best audit-PASS branch across all iterations, copy solution.py, eval.json, solve.log, plan.md to best/, write best/SELECTED.json {"iteration":"iN","branch":"bK","score":<s>}, append ledger event stage_discover. If nothing survived audits, report and stop.

Then Ablation: launch ablation-analyst (run dir). Expect best/ablations/{ablations.json, ablation.md}; ledger event stage_ablation. Ablations are secondary evidence for the paper — on ok: false, record the failure and continue; never let ablation block the pipeline. Show the user the final scoreboard (+ ablation deltas if present).

Next stage: /write-paper.

---name: discoverdescription: Run (or re-run) Stage 2 only — Discovery: Ideator plus Parallel Explore-Exploit branches (Solve → Evaluate → Audit → Rank+Select → distilled feedback), for the configured number of iterations. Requires brief.md from /investigate.---# /discover — Stage 2: Ideator + Parallel Explore-ExploitArguments: `$ARGUMENTS`Current state: !`python3 .claude/scripts/ledger.py status`Preconditions: resolve the run (`--run` or ACTIVE_RUN); `brief.md` and avalid `bibliography.jsonl` must exist — otherwise stop and point to`/investigate`. Read B and I from `run-config.json` (`--iterations` hereoverrides I; create missing `iterations/iN/branches/bK` dirs as needed).For each iteration iN:1. Launch `ideator` (run dir, iN, B, prior `distilled-feedback.md` files if any — also from previous /discover invocations). Expect `ideas.md` + `proposals/p1..pB.md`. From i2 on the ideator applies top-K retention (p1..pK refine the previous iteration's top-K surviving branches).2. Launch ALL B `solver` subagents IN ONE MESSAGE — one Agent call per branch, each given only its own `iterations/iN/branches/bK/` dir and `proposals/pK.md`. This parallel fan-out is the point of PEE; sequential launches are a protocol violation.3. When all solvers are done: launch B `evaluator` subagents in one message.4. Then B `auditor` subagents in one message.5. Rank + Select (you): read all eval.json + audit.md; audit-FAIL branches are disqualified regardless of score; write `iterations/iN/ranking.md` (table + rationale).6. Distilled feedback (you): invoke the `distill-feedback` skill (Skill tool, `skill: distill-feedback`) and write `iterations/iN/distilled-feedback.md` with exactly its section template — never improvise the format.7. Ledger: `python3 .claude/scripts/ledger.py append '{"event":"stage_iteration_iN","detail":"best=<branch>:<score>"}'`After the final iteration: select the best audit-PASS branch across alliterations, copy `solution.py`, `eval.json`, `solve.log`, `plan.md` to`best/`, write `best/SELECTED.json` `{"iteration":"iN","branch":"bK","score":<s>}`,append ledger event `stage_discover`. If nothing survived audits, report andstop.Then **Ablation**: launch `ablation-analyst` (run dir). Expect`best/ablations/{ablations.json, ablation.md}`; ledger event`stage_ablation`. Ablations are secondary evidence for the paper — on`ok: false`, record the failure and continue; never let ablation block thepipeline. Show the user the final scoreboard (+ ablation deltas if present).Next stage: `/write-paper`.

Frequently asked questions