domains / research / idea-discovery-pipeline / idea-discovery

idea-discovery

Skill card

combined

Both copies describe the same five-phase idea discovery pipeline (research-lit, idea-creator, novelty-check, research-review, research-refine-pipeline) with the same pilot constants and checkpoint concept. However, they disagree on the reviewer model (gpt-6-astra vs gpt-5.4), the AUTO_PROCEED semantics (non-blocking same-turn continuation vs proceed-after-silence), and the output directory (idea-stage/ vs project root). The wanshuiyin copy adds a RESUMABLE evidence gate with run_state.py and idea_discovery_gate.py, while the brycewang copy adds an explicit Phase 0 research brief loader and a Phase 0.5 reference paper summary procedure. Crowning one copy would hide the other's real steps and conflicting parameters, so the catalog should combine them and surface the conflicts.

Repositories

Other copies

Combined skill

# Idea Discovery Pipeline

Orchestrate a complete idea discovery workflow for: **$ARGUMENTS**

## Pipeline

```
/research-lit → /idea-creator → /novelty-check → /research-review → /research-refine-pipeline
  (survey)      (brainstorm)    (verify novel)    (critical feedback)  (refine method + plan experiments)
```

Final deliverables: a validated `IDEA_REPORT.md` with ranked ideas, plus `refine-logs/FINAL_PROPOSAL.md` and `refine-logs/EXPERIMENT_PLAN.md` for the top idea.

## Constants

- PILOT_MAX_HOURS = 2
- PILOT_TIMEOUT_HOURS = 3
- MAX_PILOT_IDEAS = 3
- MAX_TOTAL_GPU_HOURS = 8
- AUTO_PROCEED = true
- REVIEWER_MODEL = gpt-6-astra (wanshuiyin) / gpt-5.4 (brycewang) — see Conflicts
- ARXIV_DOWNLOAD = false
- COMPACT = false
- REF_PAPER = false
- OUTPUT_DIR = idea-stage/ (wanshuiyin) / project root (brycewang) — see Conflicts
- RESUMABLE = true (wanshuiyin only)

## Checkpoint Rule

Resolve AUTO_PROCEED once before Phase 0.

- AUTO_PROCEED=true (wanshuiyin): non-blocking. State the result and selected next action, then continue in the same turn. Do not end the turn at a checkpoint.
- AUTO_PROCEED=true (brycewang): present results; if the user does not respond, proceed with the best option.
- AUTO_PROCEED=false: present options, ask the user, and end the turn. Resume only after an explicit reply.

## Phase 0: Load Research Brief (brycewang)

1. Look for `RESEARCH_BRIEF.md` in the project root or the path in `$ARGUMENTS`.
2. If found, extract problem statement, constraints, prior attempts, domain knowledge, non-goals, and existing results.
3. Use the brief as primary context; merge with `$ARGUMENTS` (brief takes priority for details).
4. If no brief exists, proceed with `$ARGUMENTS` as the research direction.

## Phase 0.5: Reference Paper Summary (when REF_PAPER is set)

Skip if REF_PAPER is false.

1. If arXiv URL: invoke `/arxiv "ARXIV_ID" — download` and read the first 5 pages.
2. If local PDF: read the first 5 pages.
3. If other URL: fetch via WebFetch.
4. Generate `REF_PAPER_SUMMARY.md` with sections: Title, Authors, Venue, What They Did, Key Results, Limitations & Open Questions, Potential Improvement Directions, Codebase (if base repo is set).
5. Present a checkpoint summary to the user.

## Phase 1: Literature Survey

Invoke `/research-lit "$ARGUMENTS"`.

- Search arXiv, Google Scholar, Semantic Scholar.
- Build a landscape map: sub-directions, approaches, open problems.
- Identify structural gaps and recurring limitations.
- Output a literature summary.

Checkpoint: present the landscape summary and ask for confirmation or direction adjustment.

## Phase 2: Idea Generation

Invoke `/idea-creator` using the literature summary and any REF_PAPER_SUMMARY.md as context.

- Generate candidate ideas.
- Rank by novelty, feasibility, and impact.
- Output ranked ideas to `IDEA_REPORT.md#ranked-ideas`.

Checkpoint: present ranked ideas.

## Phase 3: Novelty Check

Invoke `/novelty-check` on the top ideas.

- Verify each idea is not already published.
- Record verification results in `IDEA_REPORT.md#novelty-verification`.

Checkpoint: present novelty results.

## Phase 4: External Critical Review

Invoke `/research-review` using REVIEWER_MODEL via Codex MCP.

- Obtain critical feedback on the top ideas.
- Record feedback in `IDEA_REPORT.md#external-critical-review`.

Checkpoint: present review feedback.

## Phase 5: Refine and Plan Experiments

Invoke `/research-refine-pipeline` on the top idea.

- Refine the method.
- Produce `refine-logs/FINAL_PROPOSAL.md` and `refine-logs/EXPERIMENT_PLAN.md`.
- Run pilot experiments within the GPU budget constants.

## Resumable Evidence Gate (wanshuiyin only)

When RESUMABLE=true:

1. Resolve `run_state.py` and `idea_discovery_gate.py` via `.aris/tools/` → `tools/` → `$ARIS_REPO/tools/` → `~/.aris/repo/tools/`.
2. Derive `<run_id>` from the direction slug and date.
3. Start the state record with `--executor <actual-model>`.
4. For each phase, mark `running` on entry and `done --artifact <path>` after the artifact is present.
5. Artifact locators:
   - research-lit: `idea-stage/IDEA_REPORT.md#literature-landscape`
   - idea-creator: `idea-stage/IDEA_REPORT.md#ranked-ideas`
   - novelty-check: `idea-stage/IDEA_REPORT.md#novelty-verification`
   - research-review: `idea-stage/IDEA_REPORT.md#external-critical-review`
   - research-refine-pipeline: `refine-logs/FINAL_PROPOSAL.md`
6. If either helper is unavailable, the final report is BLOCKED.

## Conflicts

- **REVIEWER_MODEL**: wanshuiyin/Auto-claude-code-research-in-sleep uses `gpt-6-astra`; brycewang-stanford/Auto-Empirical-Research-Skills uses `gpt-5.4`. Choose the model available in your environment.
- **AUTO_PROCEED=true semantics**: wanshuiyin treats it as non-blocking same-turn continuation; brycewang treats it as proceed-after-silence. The wanshuiyin rule is stricter and avoids ending the turn at a checkpoint.
- **OUTPUT_DIR**: wanshuiyin writes to `idea-stage/`; brycewang writes to the project root. Use `idea-stage/` for consistency with the evidence gate artifact locators.
- **RESUMABLE evidence gate**: present only in wanshuiyin. Include it if you need resumable, auditable runs; omit it for a simpler pipeline.
- **Phase 0 research brief loader and Phase 0.5 template**: present only in brycewang. Include them to support structured briefs and reference paper summaries.

## Notes

- Override defaults by appending flags, e.g., `/idea-discovery "topic" — ref paper: https://arxiv.org/abs/2406.04329` or `— compact: true`.
- When COMPACT=true, write `IDEA_CANDIDATES.md` (top 3-5 ideas) at the end of the workflow for downstream skills.
- When RENDER_HTML=true (wanshuiyin), render `IDEA_REPORT.md` to HTML via `/render-html` with `--no-review` at workflow end.