Idea Creator

Override for Codex users who want Gemini, not a second Codex agent, to act as the reviewer. Install this package after skills/skills-codex/*.

Research Idea Creator

Gemini overlay assurance: review_independence: cross-family and acceptance_status: accepted.

Generate publishable research ideas for: $ARGUMENTS

Overview

Given a broad research direction from the user, systematically generate, validate, and rank concrete research ideas. Standalone, Phase 1's landscape survey is inline (WebSearch — it does not invoke /research-lit); Phases 4-5 invoke /novelty-check, /run-experiment, and /monitor-experiment for validation and pilots. For the full sub-skill pipeline (/research-lit → idea generation → /novelty-check → /research-review), run /idea-discovery (Workflow 1), which orchestrates this skill.

Constants

• PILOT_MAX_HOURS = 2 — Skip any pilot estimated to take > 2 hours per GPU. Flag as "needs manual pilot".

• PILOT_TIMEOUT_HOURS = 3 — Hard timeout: kill pilots exceeding 3 hours. Collect partial results if available.

• MAX_PILOT_IDEAS = 3 — Pilot at most 3 ideas in parallel. Additional ideas are validated on paper only.

• MAX_TOTAL_GPU_HOURS = 8 — Total GPU budget for all pilots combined.

• REVIEWER_MODEL = gemini-review — Gemini reviewer invoked through the local gemini-review MCP bridge for brainstorming and critique. Set GEMINI_REVIEW_MODEL if you need a specific Gemini model override.

• OUTPUT_DIR = idea-stage/ — Directory for idea output files.

💡 Override via argument, e.g., /idea-creator "topic" — pilot budget: 4h per idea, 20h total.

Workflow

Phase 1: Landscape Survey (5-10 min)

Map the research area to understand what exists and where the gaps are.

1. Scan local paper library first: Check papers/ and literature/ in the project directory for existing PDFs. Read first 3 pages of relevant papers to build a baseline understanding before searching online. This avoids re-discovering what the user already knows.

2. Search recent literature using WebSearch:

   • Top venues in the last 2 years (NeurIPS, ICML, ICLR, ACL, EMNLP, etc.)
   • Recent arXiv preprints (last 6 months)
   • Use 5+ different query formulations
   • Read abstracts and introductions of the top 10-15 papers

3. Build a landscape map:

   • Group papers by sub-direction / approach
   • Identify what has been tried and what hasn't
   • Note recurring limitations mentioned in "Future Work" sections
   • Flag any open problems explicitly stated by multiple papers

4. Identify structural gaps:

   • Methods that work in domain A but haven't been tried in domain B
   • Contradictory findings between papers (opportunity for resolution)
   • Assumptions that everyone makes but nobody has tested
   • Scaling regimes that haven't been explored
   • Diagnostic questions that nobody has asked

Phase 2: Idea Generation (brainstorm with external LLM)

Use the local gemini-review MCP bridge for divergent thinking:

mcp__gemini-review__review_start:
  prompt: |
    You are a senior ML researcher brainstorming research ideas.

    Research direction: [user's direction]

    Here is the current landscape:
    [paste landscape map from Phase 1]

    Key gaps identified:
    [paste gaps from Phase 1]

    Generate 8-12 concrete research ideas. For each idea:
    1. One-sentence summary
    2. Core hypothesis (what you expect to find and why)
    3. Minimum viable experiment (what's the cheapest way to test this?)
    4. Expected contribution type: empirical finding / new method / theoretical result / diagnostic
    5. Risk level: LOW (likely works) / MEDIUM (50-50) / HIGH (speculative)
    6. Estimated effort: days / weeks / months

    Prioritize ideas that are:
    - Testable with moderate compute (8x RTX 3090 or less)
    - Likely to produce a clear positive OR negative result (both are publishable)
    - Simple at the core: one mechanism, few moving parts — an idea a colleague
      could restate after hearing it once. If the novelty only appears once a
      second module or an extra gate is added, that is packaging, not novelty.
    - Aware of the 10-15 papers above — awareness, not avoidance. Differentiation
      is the novelty check's job later, not a constraint on brainstorming.

    "Apply X to Y" is legitimate when the application would reveal something
    non-obvious — judge it by what it reveals, not by the template. A direct,
    well-executed attack on a central problem is a valid idea when nobody has
    executed it well; do not steer around crowded areas — proximity to strong
    work is a sign the problem matters, not that it is taken.

    Be genuinely creative: surprising connections, inverted assumptions,
    questions nobody thought to ask. Creativity is a new angle on a problem
    that matters — not an obscure corner nobody visits, and not extra modules
    stacked until something looks new. Generate first, filter later — the
    filters come after you, and they are strict enough. A bold, creative idea
    with a named risk beats a hedged, complicated one with none. A great idea
    is one where the answer matters regardless of which way it goes.

After this start call, immediately save the returned jobId and poll mcp__gemini-review__review_status with a bounded waitSeconds until done=true. Treat the completed status payload's response as the brainstorm output, and save the completed threadId for follow-up critique in Phase 4.

Phase 3: Mechanical consolidation + objective feasibility gate

This phase does NOT judge idea quality, novelty, or impact — those are the job of the Phase-4 cross-model reviewer (a different model family). Dropping ideas here on a same-family novelty or impact call would pre-filter the reviewer's input with same-family judgment — the opposite of why ARIS uses a cross-model reviewer at all. Phase 3 only (a) clusters near-duplicate ideas and (b) drops ideas that are OBJECTIVELY out of budget; everything else passes through ANNOTATED, not eliminated.

1. Objective feasibility gate (safe to gate here): drop an idea ONLY on a mechanical, budget-based fact — estimated compute > 1 week of available GPU time, OR a dataset that is provably unavailable. Do NOT drop on "implementation looks complex" — annotate complexity instead.

2. Novelty signal — ANNOTATE, do not eliminate: do 2-3 targeted searches and attach a prior_work note (what looks related, with links). This is input for the Phase-4 reviewer, not a filter; full /novelty-check runs in Phase 4. Do NOT drop an idea here because it "might already be done."

3. Impact signal — ANNOTATE, do not eliminate: attach a one-line so_what note (why the result would matter either way). Do NOT drop on a same-family "a reviewer wouldn't care" call — that is exactly what the Phase-4 cross-model reviewer is for.

Every feasible, non-duplicate idea — with its prior_work and so_what annotations — proceeds to Phase 4, where the cross-model reviewer does the quality/novelty narrowing.

Phase 4: Deep Validation (for top ideas)

For each surviving idea, run a deeper evaluation:

1. Novelty check: Use the /novelty-check workflow (multi-source search + Gemini cross-verification) for each idea

2. Critical review: Use mcp__gemini-review__review_reply_start with the saved completed threadId:

   mcp__gemini-review__review_reply_start:
     threadId: [saved completed threadId from Phase 2]
     prompt: |
       Here are our top ideas after filtering:
       [paste surviving ideas with novelty check results]
   
       For each, make the strongest case both ways:
       - What is the best case FOR it — what would make this the paper people cite?
       - What's the strongest objection a reviewer would raise?
       - What's the most likely failure mode?
       - Rank by expected information and upside within the pilot budget — which results would matter most, whichever way they come out?
       - Which 2-3 would you actually work on?
   
       Rank; do not rewrite. An objection is answered or recorded as a named
       risk on the idea — never absorbed by adding a module, a gate, or a
       qualifier. A bold idea with a named risk outranks a hedged idea with
       none, and complexity added since the brainstorm is a red flag, not
       progress. And do not let your picks be uniformly the safest — if
       the top set is all LOW-risk, name the high-upside idea that most
       deserves a pilot slot and what result would convince you.
   

   After this start call, immediately save the returned jobId and poll mcp__gemini-review__review_status with a bounded waitSeconds until done=true. Treat the completed status payload's response as the follow-up critique.

3. Combine rankings: Merge your assessment with Gemini's ranking. Select top 2-3 ideas for pilot experiments.

Phase 5: Parallel Pilot Experiments (for top 2-3 ideas)

Before committing to a full research effort, run cheap pilot experiments to get empirical signal. This is the key differentiator from paper-only validation.

1. Design pilots: For each top idea, define the minimal experiment that would give a positive or negative signal:

   • Single seed, small scale (e.g., small dataset subset, fewer epochs)
   • Target: 30 min - PILOT_MAX_HOURS per pilot on 1 GPU
   • Estimate GPU-hours BEFORE launching. If estimated time > PILOT_MAX_HOURS, reduce scale (fewer epochs, smaller subset) or flag as "needs manual pilot"
   • Decision criterion defined upfront — including what a positive, negative, and null outcome would each teach. Metric improvement is not required for a diagnostic contribution.

2. Deploy in parallel: Use /run-experiment to launch pilots on different GPUs simultaneously:

   GPU 0: Pilot for Idea 1
   GPU 1: Pilot for Idea 2
   GPU 2: Pilot for Idea 3
   

   Use run_in_background: true to launch all at once.

3. Collect results: Use /monitor-experiment to check progress. If any pilot exceeds PILOT_TIMEOUT_HOURS, kill it and collect partial results. Once all pilots complete (or timeout), compare:

   • Which ideas showed positive signal?
   • Which showed null/negative results? Classify each: core-hypothesis refuted, informative negative (often publishable), or underpowered pilot — do not eliminate by sign alone.
   • Any surprising findings that suggest a pivot?
   • Total GPU-hours consumed (track against MAX_TOTAL_GPU_HOURS budget)

4. Re-rank based on empirical evidence: Update the idea ranking using pilot results. An idea with strong pilot signal jumps ahead of a theoretically appealing but untested idea.

Note: Skip this phase if the ideas are purely theoretical or if no GPU is available. Flag skipped ideas as "needs pilot validation" in the report.

Phase 6: Output — Ranked Idea Report

Write a structured report to idea-stage/IDEA_REPORT.md:

Lead every recommended idea with its method, in plain language. Before any hypothesis, novelty score, or claim, state in 2–4 concrete steps what we actually build / train / run — no jargon, no claim-IDs. The reader must understand what we do before what we claim; claims (hypothesis, validation, expected outcome) come after and read as the method's acceptance criteria.

# Research Idea Report

**Direction**: [user's research direction]
**Generated**: [date]
**Ideas evaluated**: X generated → Y survived filtering → Z piloted → W recommended

## Landscape Summary
[3-5 paragraphs on the current state of the field]

## Recommended Ideas (ranked)

### Idea 1: [title]
- **Method (what we actually do)**: [2–4 concrete steps in plain language — what we build / train / run. No jargon, no claim-IDs, no hypothesis yet. Lead with this so the reader grasps the approach first.]
- **Hypothesis**: [one sentence]
- **Minimum experiment**: [concrete description]
- **Expected outcome**: [what success/failure looks like]
- **Novelty**: X/10 — closest work: [paper]
- **Feasibility**: [compute, data, implementation estimates]
- **Risk**: LOW/MEDIUM/HIGH
- **Contribution type**: empirical / method / theory / diagnostic
- **Pilot result**: [POSITIVE: metric +X% / NEGATIVE: no signal / SKIPPED: needs GPU]
- **Reviewer's likely objection**: [strongest counterargument]
- **Why we should do this**: [1-2 sentences]

### Idea 2: [title]
...

## Eliminated Ideas (for reference)
| Idea | Reason eliminated |
|------|-------------------|
| ... | Already done by [paper] |
| ... | Requires > 1 week GPU time |
| ... | Result wouldn't be interesting either way |

## Pilot Experiment Results
| Idea | GPU | Time | Key Metric | Signal |
|------|-----|------|------------|--------|
| Idea 1 | GPU 0 | 45 min | +2.3% CE | POSITIVE |
| Idea 2 | GPU 1 | 30 min | -0.1% CE | NEGATIVE |
| Idea 3 | GPU 2 | 1.5 hr | +0.8% CE | WEAK POSITIVE |

## Suggested Execution Order
1. Start with Idea 1 (highest decision value after the pilot)
2. Idea 3 as backup (weak signal, may need larger scale to confirm)
3. Idea 2 eliminated by pilot — negative result documented

## Next Steps
- [ ] Scale up Idea 1 to full experiment (multi-seed, full dataset)
- [ ] If confirmed, invoke /auto-review-loop for full iteration

Output Protocols

Follow these shared protocols for all output files:

• Output Versioning Protocol — write timestamped file first, then copy to fixed name
• Output Manifest Protocol — log every output to MANIFEST.md
• Output Language Protocol — respect the project's language setting

Key Rules

• Large file handling: If the Write tool fails due to file size, immediately retry using Bash (cat << 'EOF' > file) to write in chunks. Do NOT ask the user for permission — just do it silently.

• The user provides a DIRECTION, not an idea. Your job is to generate the ideas.

• Quantity first, quality second: brainstorm broadly, then narrow only to allocate pilot budget — annotate the rest, don't paper-kill them.

• A good negative result is just as publishable as a positive one. Prioritize ideas where the answer matters regardless of direction.

• Don't fall in love with any idea before validating it — but let evidence do the killing, not anticipated objections.

• Always estimate compute cost. An idea that needs 1000 GPU-hours is not actionable for most researchers.

• "Apply X to Y" is legitimate when Y can reveal a non-obvious interaction, failure mode, or finding — judge the revelation, not the template.

• Include eliminated ideas in the report — they save future time by documenting dead ends.

• If the user's direction is broad (e.g., "NLP"), use Phase 1 to derive 2-3 concrete frames and generate across them — ask the user only when a missing constraint would materially change the pilot slate. A good direction is 1-2 sentences specifying the problem, domain, and constraint — e.g., "factorized gap in discrete diffusion LMs" or "sample efficiency of offline RL with image observations". Without sufficient specificity, generated ideas will be too vague to run experiments on.

Composing with Other Skills

After this skill produces the ranked report:

/idea-creator "direction"     → ranked ideas
/novelty-check "top idea"     → deep novelty verification (already done in Phase 4, but user can re-run)
/research-review "top idea"   → external critical feedback
implement                     → write code
/run-experiment               → deploy to GPU
/auto-review-loop             → iterate until submission-ready