idea-pipeline

name	idea-pipeline
description	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.
argument-hint	[research-direction] [-- venue: ICML|VLDB|NeurIPS]
allowed-tools	Bash(*), Read, Write, Edit, Grep, Glob, WebSearch, WebFetch, Agent, Skill, mcp__codex__codex, mcp__codex__codex-reply, mcp__zotero__*, mcp__obsidian-vault__*

Idea Pipeline: Full Idea Discovery Orchestrator

Orchestrate a complete idea discovery workflow for: $ARGUMENTS

Constants

• REFINE_TOP_N = 2 — Number of top ideas to refine in Phase 4.
• REVIEWER_MODEL = gpt-5.4 — Passed through to sub-skills. Must be an OpenAI model (e.g., gpt-5.4, o3, gpt-4o).
• DEFAULT_VENUE = ICML — Default venue for screening when -- venue: is not specified.

Override defaults by telling the skill, e.g., /idea-pipeline "topic" -- venue: NeurIPS.

Autonomous Operation

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.

State Persistence (PIPELINE_STATE.json)

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:

• If exists AND direction matches $ARGUMENTS → resume from the next incomplete phase (skip already-completed phases).
• If exists but direction differs → start fresh (rename old state file to outputs/PIPELINE_STATE.json.bak).
• If not exists → start fresh.

Update current_phase, completed_phases, and phase_timestamps after each phase completes, using the Write tool.

Overview

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.

Pre-Flight Check

Do NOT probe or test any tools before starting. Proceed directly to Phase 1.

Pipeline Detail

Phase 1: Literature Survey

Invoke /lit-survey "$ARGUMENTS" via the Skill tool.

What this does:

• Searches arXiv, Google Scholar, Semantic Scholar, local PDFs, Zotero, Obsidian
• Builds landscape map with gap identification matrix
• Outputs: outputs/LANDSCAPE.md, outputs/LANDSCAPE.json

Checkpoint 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.

Phase 2: Idea Generation + Filtering

Invoke /idea-gen "$ARGUMENTS" via the Skill tool.

What this does:

• Reads outputs/LANDSCAPE.json from Phase 1
• Brainstorms 8-12 ideas via external LLM (xhigh reasoning)
• Filters by feasibility, novelty quick-check, impact
• Applies Prof. He's 4-dimension filter (threshold: 12/20)
• Runs anti-pattern check
• Outputs: outputs/IDEAS_RAW.md, outputs/IDEAS_FILTERED.md

Checkpoint 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.

Phase 3: Multi-Dimensional Screening

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:

• Module A: Novelty assessment (multi-source search + cross-model verification)
• Module B: Venue reviewer simulation (3 reviewers + meta review)
• Module C: Strategic fit assessment (5 dimensions)
• Composite scoring and ranking
• Outputs: outputs/SCREENING_REPORT.md, outputs/SCREENING_RANKED.md

Checkpoint 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.

Phase 4: Deep Refinement

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:

• Freezes Problem Anchor
• Extracts logical skeleton
• Iteratively refines via external LLM (up to 5 rounds)
• 7-dimension scoring, threshold 9/10
• Outputs: refine-logs/FINAL_PROPOSAL.md, refine-logs/REFINEMENT_REPORT.md

Checkpoint 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.

Phase 5: Final Report

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

Writing the Final Report

1. Ensure directories exist: mkdir -p outputs/ refine-logs/
2. Read output files from each phase to extract summaries and data.
3. Write outputs/IDEA_DISCOVERY_REPORT.md using the Write tool.
4. Large file fallback: If Write fails due to file size, immediately retry using Bash:

   cat << 'REPORT_EOF' > outputs/IDEA_DISCOVERY_REPORT.md
   [content]
   REPORT_EOF
   

   Do NOT ask the user for permission — just do it silently.

Key Rules

• Log decisions, never ask. All checkpoint summaries go to outputs/PIPELINE_LOG.md. Never ask the user questions, present options, or wait for confirmation.
• If a sub-skill fails, log the error and continue with degraded quality. Append the error details to outputs/PIPELINE_LOG.md and proceed to the next phase with whatever data is available. Do not stop the pipeline.
• Always produce a final report, even if some phases failed. Mark failed phases clearly in the report with [PHASE FAILED] and include the error details.
• 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.
• Kill ideas early. It's better to kill 10 bad ideas in Phase 3 than to refine one and fail.
• Document everything. Dead ends are valuable for future reference. Eliminated ideas should always include the reason for elimination.
• Pass venue parameter through. Parse -- venue: from $ARGUMENTS and forward to /idea-screen. If not specified, use DEFAULT_VENUE.
• Update PIPELINE_STATE.json after every phase. This enables resumption if the pipeline is interrupted.

Composing

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)