Full idea discovery pipeline: literature survey → idea generation → multi-dimensional screening → deep refinement. Use when user says "find ideas", "idea discovery", "full pipeline", "从零开始找方向", "找idea全流程", or wants the complete workflow from research direction to refined proposal.
Orchestrate a complete idea discovery workflow for: $ARGUMENTS
gpt-5.4 — Passed through to sub-skills. Must be an OpenAI model (e.g., gpt-5.4, o3, gpt-4o).-- venue: is not specified.Override defaults by telling the skill, e.g.,
/idea-pipeline "topic" -- venue: NeurIPS.
This pipeline runs fully autonomously once invoked. No user interaction is required.
Logging: All checkpoint decisions are appended to outputs/PIPELINE_LOG.md with timestamps. Review this file to understand the pipeline's autonomous decisions.
Error handling: If any phase fails, the error is logged and the pipeline attempts graceful degradation (see individual skill documentation). The pipeline will NOT stop to ask the user.
After EACH phase completes, write/update outputs/PIPELINE_STATE.json with the current pipeline state:
{
"direction": "$ARGUMENTS",
"venue": "[venue]",
"current_phase": 2,
"completed_phases": [1],
"started_at": "[ISO timestamp]",
"phase_timestamps": {"1_completed": "[ISO timestamp]"}
}
At pipeline START, check if outputs/PIPELINE_STATE.json exists:
direction matches $ARGUMENTS → resume from the next incomplete phase (skip already-completed phases).direction differs → start fresh (rename old state file to outputs/PIPELINE_STATE.json.bak).Update current_phase, completed_phases, and phase_timestamps after each phase completes, using the Write tool.
This skill chains all 4 sub-skills into a single automated pipeline:
/idea-pipeline "research direction"
Phase 1: /lit-survey → landscape map + gaps
Phase 2: /idea-gen → 8-12 ideas, filtered to 4-6
Phase 3: /idea-screen → multi-dimensional screening + ranking
Phase 4: /idea-refine → refined proposal for top ideas
Phase 5: Final report aggregation
Each phase builds on the previous one's output. Checkpoint results are logged to outputs/PIPELINE_LOG.md and the pipeline auto-continues without user interaction.
Do NOT probe or test any tools before starting. Proceed directly to Phase 1.
Invoke /lit-survey "$ARGUMENTS" via the Skill tool.
What this does:
outputs/LANDSCAPE.md, outputs/LANDSCAPE.jsonCheckpoint 1 — Literature Survey (auto-logged):
Append to outputs/PIPELINE_LOG.md:
## [Timestamp] Phase 1 Complete: Literature Survey
- Found N papers, identified M research gaps
- Top themes: [list top 3 themes]
- Gap matrix written to outputs/LANDSCAPE.md
- **Auto-decision**: Proceeding to Phase 2 (Idea Generation) with all identified gaps
Then proceed immediately to the next phase. Do NOT ask the user any questions. Do NOT present options. Do NOT wait for confirmation.
Invoke /idea-gen "$ARGUMENTS" via the Skill tool.
What this does:
outputs/LANDSCAPE.json from Phase 1outputs/IDEAS_RAW.md, outputs/IDEAS_FILTERED.mdCheckpoint 2 — Idea Generation (auto-logged):
Append to outputs/PIPELINE_LOG.md:
## [Timestamp] Phase 2 Complete: Idea Generation
- Generated X ideas, filtered to Y after He filter (threshold 12/20)
- Top ideas: [list titles with He scores]
- **Auto-decision**: Screening ALL Y filtered ideas in Phase 3
Then proceed immediately to the next phase. Do NOT ask the user any questions. Do NOT present options. Do NOT wait for confirmation.
Invoke /idea-screen for ALL filtered ideas, with the venue parameter.
Parse -- venue: from the original $ARGUMENTS and pass through. If not specified, use DEFAULT_VENUE.
What this does:
outputs/SCREENING_REPORT.md, outputs/SCREENING_RANKED.mdCheckpoint 3 — Idea Screening (auto-logged):
Append to outputs/PIPELINE_LOG.md:
## [Timestamp] Phase 3 Complete: Idea Screening
- Screened Y ideas for venue [VENUE]
- Ranking: [list top ideas with composite scores]
- Recommendations: [PROCEED/REVISE/REJECT counts]
- **Auto-decision**: Refining top REFINE_TOP_N ideas in Phase 4
Then proceed immediately to the next phase. Do NOT ask the user any questions. Do NOT present options. Do NOT wait for confirmation.
For the top REFINE_TOP_N ideas (by composite score from Phase 3):
Invoke /idea-refine "[idea description + screening results]" for each.
What this does:
refine-logs/FINAL_PROPOSAL.md, refine-logs/REFINEMENT_REPORT.mdCheckpoint 4 — Idea Refinement (auto-logged):
Append to outputs/PIPELINE_LOG.md:
## [Timestamp] Phase 4 Complete: Idea Refinement
- Refined REFINE_TOP_N ideas
- Idea 1: [title] — Score X/10, Verdict: READY/REVISE/RETHINK, Rounds: N/5
- Idea 2: [title] — Score X/10, Verdict: READY/REVISE/RETHINK, Rounds: N/5
- **Auto-decision**: Generating final report in Phase 5
Then proceed immediately to the next phase. Do NOT ask the user any questions. Do NOT present options. Do NOT wait for confirmation.
Aggregate all outputs into outputs/IDEA_DISCOVERY_REPORT.md:
# Idea Discovery Report
**Direction**: $ARGUMENTS
**Date**: [today]
**Venue**: [venue]
**Pipeline**: lit-survey → idea-gen → idea-screen → idea-refine
## Executive Summary
[2-3 sentences: best idea, composite score, refinement verdict, recommended next step]
## Literature Landscape
[summary from Phase 1, with link to full outputs/LANDSCAPE.md]
## Ideas Generated and Filtered
| # | Title | He Score | Feasibility | Status |
|---|-------|----------|-------------|--------|
[all ideas from Phase 2 with their fate]
## Screening Results (venue: [VENUE])
| Rank | Idea | Novelty | Venue | Strategic | Composite | Recommendation |
|------|------|---------|-------|-----------|-----------|----------------|
[ranking table from Phase 3]
## Refined Proposals
### Idea 1: [title] — RECOMMENDED
- Composite score: X.X/10
- Refinement verdict: READY/REVISE (score: X/10, N rounds)
- Method thesis: [one sentence]
- Key strength: [from screening]
- Key risk: [from screening]
- Proposal: `refine-logs/FINAL_PROPOSAL.md`
### Idea 2: [title] — BACKUP
[similar structure]
## Eliminated Ideas
| Idea | Stage | Reason |
|------|-------|--------|
| ... | Phase 2 (feasibility) | Requires unavailable dataset |
| ... | Phase 2 (He filter) | Score 10/20 |
| ... | Phase 3 (novelty) | Already published by [paper] |
| ... | Phase 3 (venue) | All reviewers: Reject |
## All Output Files
- `outputs/LANDSCAPE.md` — Literature landscape
- `outputs/LANDSCAPE.json` — Structured landscape data
- `outputs/IDEAS_RAW.md` — All generated ideas
- `outputs/IDEAS_FILTERED.md` — Filtered ideas
- `outputs/SCREENING_REPORT.md` — Full screening report
- `outputs/SCREENING_RANKED.md` — Ranked results
- `refine-logs/FINAL_PROPOSAL.md` — Refined proposal
- `refine-logs/REFINEMENT_REPORT.md` — Refinement process
- `outputs/IDEA_DISCOVERY_REPORT.md` — This report
- `outputs/PIPELINE_LOG.md` — Autonomous decision log
- `outputs/PIPELINE_STATE.json` — Pipeline state for resumption
mkdir -p outputs/ refine-logs/outputs/IDEA_DISCOVERY_REPORT.md using the Write tool.cat << 'REPORT_EOF' > outputs/IDEA_DISCOVERY_REPORT.md
[content]
REPORT_EOF
Do NOT ask the user for permission — just do it silently.outputs/PIPELINE_LOG.md. Never ask the user questions, present options, or wait for confirmation.outputs/PIPELINE_LOG.md and proceed to the next phase with whatever data is available. Do not stop the pipeline.[PHASE FAILED] and include the error details.cat << 'EOF' > file) to write in chunks. Do NOT ask the user for permission — just do it silently.-- venue: from $ARGUMENTS and forward to /idea-screen. If not specified, use DEFAULT_VENUE.This is the top-level workflow. Individual skills can also be used standalone:
/lit-survey "topic" — just the literature survey
/idea-gen "direction" — just idea generation
/idea-screen "ideas" — just screening
/idea-refine "idea" — just refinement
/idea-pipeline "direction" ← full pipeline (this skill)