/idea-discovery-pipeline

name: idea-discovery-pipeline
description: The full pipeline for idea generation. It generates 8-12 novel research ideas from literature gaps and evaluates each on novelty, feasibility, and domain fit. Orchestrates lit-review → generate-idea → novelty-check → idea-review → experiment-design-pipeline to go from a broad research direction to a validated, pilot-tested idea with a refined proposal and experiment plan. Produces output/IDEA_REPORT.md plus refinement and experiment artifacts.
argument-hint: [research-direction]
allowed-tools: Bash(*), Read, Write, Edit, Grep, Glob, WebSearch, WebFetch, Agent, Skill

Skill: idea-discovery-pipeline

You are the orchestrator for a complete idea discovery workflow for: **$ARGUMENTS**

---

Overview

This skill chains sub-skills into a single automated pipeline:

/lit-review → /generate-idea → /novelty-check → /idea-review → /experiment-design-pipeline

Each phase builds on the previous one's output. The final deliverables are:

• `output/LIT_REVIEW_REPORT.md` — literature landscape and ranked gaps
• `output/IDEA_REPORT.md` — ranked, validated, pilot-tested ideas (the top-level report)
• `output/NOVELTY_REPORT.md` — deep novelty verdicts for top ideas
• `output/IDEA_REVIEW_REPORT.md` — external critical review of the top idea
• `output/refine-logs/FINAL_PROPOSAL.md` + `output/refine-logs/REFINE_REPORT.md` — refined proposal
• `output/EXPERIMENT_PLAN.md` + `output/EXPERIMENT_TRACKER.md` — experiment roadmap
• `output/EXP_PIPELINE_SUMMARY.md` — pipeline summary from experiment-design-pipeline

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. If not configured properly, use Claude Code subagent.
• **ARXIV_DOWNLOAD = false** — When `true`, `/lit-review` downloads the top relevant arXiv PDFs during Phase 1. When `false` (default), only fetches metadata. Passed through to `/lit-review`.
• **COMPACT = false** — When `true`, generate compact summary files for short-context models and session recovery. Writes `output/IDEA_CANDIDATES.md` (top 3-5 ideas only) at the end of this workflow. Downstream skills read this instead of the full `output/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 (`output/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-pipeline "topic" — ref paper: https://arxiv.org/abs/2406.04329` or `/idea-discovery-pipeline "topic" — compact: true`.

Pipeline

Phase 0: Load Context and Research Plan

At the very start, read the most relevant existing files if they exist (use `Read`; skip silently if missing). This mirrors the Phase 0 pattern used by each subskill and ensures the pipeline never re-does stable work:

• `CLAUDE.md` — project dashboard, control flags, canonical output paths
• `handoff.json` — if present, check `pipeline.stage` and `recovery.resume_skill`
• `RESEARCH_PLAN.md` in the project root (or path passed as `$ARGUMENTS`) — detailed research brief
• `output/LIT_REVIEW_REPORT.md` — prior literature review, if any
• `output/IDEA_REPORT.md` — prior idea candidates, if any
• `output/NOVELTY_REPORT.md` — prior novelty verdicts, if any
• `output/IDEA_REVIEW_REPORT.md` — prior external review, if any
• `output/REF_PAPER_SUMMARY.md` — prior reference paper summary, if any
• `output/refine-logs/FINAL_PROPOSAL.md` — prior refined proposal, if any

If `RESEARCH_PLAN.md` is found, 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_PLAN.md` and a one-line `$ARGUMENTS` exist, merge them (the plan takes priority for details; the argument sets the direction). If no plan exists, proceed with `$ARGUMENTS` as the research direction.

Decide whether to resume from a prior stage or re-run:

• If a prior artifact is stale or materially different from the current request, re-run that phase.
• If it is strong and aligned, reuse it and jump to the next phase.
• When in doubt, prefer re-running over planning on top of the wrong idea.

> 💡 Create a plan from the template: `cp templates/RESEARCH_PLAN_TEMPLATE.md RESEARCH_PLAN.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:

1. **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)

2. **If local PDF path** (e.g., `papers/reference.pdf`):

• Read the PDF directly (first 5 pages)

3. **If other URL**:

• Fetch and extract content via WebFetch

4. **Generate `output/REF_PAPER_SUMMARY.md`** (canonical path — use this everywhere):

# Reference Paper Summary