Each phase builds on the previous one's output. The final deliverables are a validated IDEA_REPORT.md with ranked ideas, plus a refined proposal (refine-logs/FINAL_PROPOSAL.md) and experiment plan (refine-logs/EXPERIMENT_PLAN.md) for the top idea.
Constants
PILOT_MAX_HOURS = 2 — Skip any pilot experiment estimated to take > 2 hours per GPU. Flag as "needs manual pilot" in the report.
PILOT_TIMEOUT_HOURS = 3 — Hard timeout: kill any running pilot that exceeds 3 hours. Collect partial results if available.
MAX_PILOT_IDEAS = 3 — Run pilots for at most 3 top ideas in parallel. Additional ideas are validated on paper only.
MAX_TOTAL_GPU_HOURS = 8 — Total GPU budget across all pilots. If exceeded, skip remaining pilots and note in report.
AUTO_PROCEED = true — If user doesn't respond at a checkpoint, automatically proceed with the best option after presenting results. Set to false to always wait for explicit user confirmation.
REVIEWER_MODEL = gpt-5.4 — Model used via Codex MCP. Must be an OpenAI model (e.g., gpt-5.4, o3, gpt-4o). Passed to sub-skills.
ARXIV_DOWNLOAD = false — When true, /research-lit downloads the top relevant arXiv PDFs during Phase 1. When false (default), only fetches metadata. Passed through to /research-lit.
COMPACT = false — When true, generate compact summary files for short-context models and session recovery. Writes IDEA_CANDIDATES.md (top 3-5 ideas only) at the end of this workflow. Downstream skills read this instead of the full IDEA_REPORT.md.
REF_PAPER = false — Reference paper to base ideas on. Accepts: local PDF path, arXiv URL, or any paper URL. When set, the paper is summarized first (REF_PAPER_SUMMARY.md), then idea generation uses it as context. Combine with base repo for "improve this paper with this codebase" workflows.
💡 These are defaults. Override by telling the skill, e.g., /idea-discovery "topic" — ref paper: https://arxiv.org/abs/2406.04329 or /idea-discovery "topic" — compact: true.
Pipeline
Phase 0: Load Research Brief (if available)
Before starting any other phase, check for a detailed research brief in the project:
Look for RESEARCH_BRIEF.md in the project root (or path passed as $ARGUMENTS)
If found, read it and extract:
Problem statement and context
Constraints (compute, data, timeline, venue)
What the user already tried / what didn't work
Domain knowledge and non-goals
Existing results (if any)
Use this as the primary context for all subsequent phases — it replaces the one-line prompt
If both RESEARCH_BRIEF.md and a one-line $ARGUMENTS exist, merge them (brief takes priority for details, argument sets the direction)
If no brief exists, proceed normally with $ARGUMENTS as the research direction.
💡 Create a brief from the template: cp templates/RESEARCH_BRIEF_TEMPLATE.md RESEARCH_BRIEF.md
Phase 0.5: Reference Paper Summary (when REF_PAPER is set)
Skip entirely if REF_PAPER is false.
Summarize the reference paper before searching the literature:
If arXiv URL (e.g., https://arxiv.org/abs/2406.04329):
Invoke /arxiv "ARXIV_ID" — download to fetch the PDF
Read the first 5 pages (title, abstract, intro, method overview)
If local PDF path (e.g., papers/reference.pdf):
Read the PDF directly (first 5 pages)
If other URL:
Fetch and extract content via WebFetch
Generate REF_PAPER_SUMMARY.md:
# Reference Paper Summary
**Title**: [paper title]
**Authors**: [authors]
**Venue**: [venue, year]
## What They Did
[2-3 sentences: core method and contribution]
## Key Results
[Main quantitative findings]
## Limitations & Open Questions
[What the paper didn't solve, acknowledged weaknesses, future work suggestions]
## Potential Improvement Directions
[Based on the limitations, what could be improved or extended?]
## Codebase
[If `base repo` is also set: link to the repo and note which parts correspond to the paper]
🚦 Checkpoint: Present the summary to the user:
📄 Reference paper summarized:
- Title: [title]
- Key limitation: [main gap]
- Improvement directions: [2-3 bullets]
Proceeding to literature survey with this as context.
Phase 1 and Phase 2 will use REF_PAPER_SUMMARY.md as additional context — /research-lit searches for related and competing work, /idea-creator generates ideas that build on or improve the reference paper.
Phase 1: Literature Survey
Invoke /research-lit to map the research landscape:
/research-lit "$ARGUMENTS"
What this does:
Search arXiv, Google Scholar, Semantic Scholar for recent papers
Build a landscape map: sub-directions, approaches, open problems
Identify structural gaps and recurring limitations
Output a literature summary (saved to working notes)
🚦 Checkpoint: Present the landscape summary to the user. Ask:
📚 Literature survey complete. Here's what I found:
- [key findings, gaps, open problems]
Does this match your understanding? Should I adjust the scope before generating ideas?
(If no response, I'll proceed with the top-ranked direction.)
User approves (or no response + AUTO_PROCEED=true) → proceed to Phase 2 with best direction.
User requests changes (e.g., "focus more on X", "ignore Y", "too broad") → refine the search with updated queries, re-run /research-lit with adjusted scope, and present again. Repeat until the user is satisfied.
Phase 2: Idea Generation + Filtering + Pilots
Invoke /idea-creator with the landscape context (and REF_PAPER_SUMMARY.md if available):
/idea-creator "$ARGUMENTS"
What this does:
If REF_PAPER_SUMMARY.md exists, include it as context — ideas should build on, improve, or extend the reference paper
Brainstorm 8-12 concrete ideas via GPT-5.4 xhigh
Filter by feasibility, compute cost, quick novelty search
Deep validate top ideas (full novelty check + devil's advocate)
Run parallel pilot experiments on available GPUs (top 2-3 ideas)
Rank by empirical signal
Output IDEA_REPORT.md
🚦 Checkpoint: Present IDEA_REPORT.md ranked ideas to the user. Ask:
💡 Generated X ideas, filtered to Y, piloted Z. Top results:
1. [Idea 1] — Pilot: POSITIVE (+X%)
2. [Idea 2] — Pilot: WEAK POSITIVE (+Y%)
3. [Idea 3] — Pilot: NEGATIVE, eliminated
Which ideas should I validate further? Or should I regenerate with different constraints?
(If no response, I'll proceed with the top-ranked ideas.)
User picks ideas (or no response + AUTO_PROCEED=true) → proceed to Phase 3 with top-ranked ideas.
User unhappy with all ideas → collect feedback ("what's missing?", "what direction do you prefer?"), update the prompt with user's constraints, and re-run Phase 2 (idea generation). Repeat until the user selects at least 1 idea.
User wants to adjust scope → go back to Phase 1 with refined direction.
Phase 3: Deep Novelty Verification
For each top idea (positive pilot signal), run a thorough novelty check:
/novelty-check "[top idea 1 description]"
/novelty-check "[top idea 2 description]"
What this does:
Multi-source literature search (arXiv, Scholar, Semantic Scholar)
Cross-verify with GPT-5.4 xhigh
Check for concurrent work (last 3-6 months)
Identify closest existing work and differentiation points
Update IDEA_REPORT.md with deep novelty results. Eliminate any idea that turns out to be already published.
Phase 4: External Critical Review
For the surviving top idea(s), get brutal feedback:
/research-review "[top idea with hypothesis + pilot results]"
What this does:
GPT-5.4 xhigh acts as a senior reviewer (NeurIPS/ICML level)
Scores the idea, identifies weaknesses, suggests minimum viable improvements
Provides concrete feedback on experimental design
Update IDEA_REPORT.md with reviewer feedback and revised plan.
Write IDEA_CANDIDATES.md — a lean summary of the top 3-5 surviving ideas:
# Idea Candidates
| # | Idea | Pilot Signal | Novelty | Reviewer Score | Status |
|---|------|-------------|---------|---------------|--------|
| 1 | [title] | +X% | Confirmed | X/10 | RECOMMENDED |
| 2 | [title] | +Y% | Confirmed | X/10 | BACKUP |
| 3 | [title] | Negative | — | — | ELIMINATED |
## Active Idea: #1 — [title]
- Hypothesis: [one sentence]
- Key evidence: [pilot result]
- Next step: /experiment-bridge or /research-refine
This file is intentionally small (~30 lines) so downstream skills and session recovery can read it without loading the full IDEA_REPORT.md (~200+ lines).
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.
Don't skip phases. Each phase filters and validates — skipping leads to wasted effort later.
Checkpoint between phases. Briefly summarize what was found before moving on.
Kill ideas early. It's better to kill 10 bad ideas in Phase 3 than to implement one and fail.
Empirical signal > theoretical appeal. An idea with a positive pilot outranks a "sounds great" idea without evidence.
Document everything. Dead ends are just as valuable as successes for future reference.
Be honest with the reviewer. Include negative results and failed pilots in the review prompt.
Feishu notifications are optional. If ~/.claude/feishu.json exists, send checkpoint at each phase transition and pipeline_done at final report. If absent/off, skip silently.
Composing with Workflow 2
After this pipeline produces a validated top idea:
/idea-discovery "direction" ← you are here (Workflow 1, includes method refinement + experiment planning)
/run-experiment ← deploy experiments from the plan
/auto-review-loop "top idea" ← Workflow 2: iterate until submission-ready
Or use /research-pipeline for the full end-to-end flow.
