drafting-research-plan-from-idea-folder

name	drafting-research-plan-from-idea-folder
description	Use when given a specific Claude Code paper-idea folder containing `outputs/IDEA_DISCOVERY_REPORT.md`, `outputs/IDEAS_RAW.md`, `outputs/IDEAS_FILTERED.md`, and `refine-logs/FINAL_PROPOSAL.md`, and asked to produce a publication-grade `work.md` for the strongest idea, with idea audit, rigorous theory development, and resource-bounded experiment planning for a future ML paper.

Drafting Research Plan From Idea Folder

Overview

Turn one specified folder into one publication-grade work.md. This skill is the single entry point for the whole pipeline. Keep one user-facing skill, but execute the work in strict stage order.

Use <target-folder> as the input variable.

Required Inputs

Required:

• <target-folder>/outputs/IDEA_DISCOVERY_REPORT.md
• <target-folder>/outputs/IDEAS_RAW.md
• <target-folder>/outputs/IDEAS_FILTERED.md
• <target-folder>/refine-logs/FINAL_PROPOSAL.md

Optional but preferred:

• <target-folder>/refine-logs/FINAL_PROPOSAL_IDEAxx.md for the selected idea

Source Priority

Use IDEA_DISCOVERY_REPORT.md to choose the best idea and the final recommendation.

Use FINAL_PROPOSAL.md and any matching FINAL_PROPOSAL_IDEAxx.md to recover technical detail, theory hooks, and experiment hooks.

Use IDEAS_FILTERED.md and IDEAS_RAW.md only for supplementary context. If files disagree, trust IDEA_DISCOVERY_REPORT.md and FINAL_PROPOSAL.md over filtered or raw files. Do not invent a new ranking.

Execution Order

1. Load references/stage-1-idea-audit.md and complete the idea audit.
2. Load references/stage-2-theory-construction.md and build the theory section from the Stage 1 contract.
3. Load references/stage-3-experiment-program.md and build the experiment section from the Stage 1 claims and Stage 2 theory.
4. Load references/output-contract.md and write <target-folder>/work.md.

Do not skip stages. Do not treat the stages as independent user-visible skills. The output of each stage is the input contract for the next one.

Gating Rules

Stage 1 must end with:

• one chosen idea
• one-sentence paper thesis
• one claims inventory
• one proceed status: PROCEED, PROCEED WITH MAJOR RISK, or NO-GO

If Stage 1 reveals the strongest idea is still weak, continue honestly. Keep the best available version, but mark the risks explicitly instead of writing as if the idea is already top-tier ready.

Stage 2 must distinguish among:

• fully derived results
• careful proof sketches
• open proof obligations

Never upgrade a sketch into a proof by wording alone.

Stage 3 must fit the resource budget. If the full plan exceeds budget, cut to an MVP experiment set and explain what was deferred.

Output

Write exactly one final file:

• <target-folder>/work.md

The final document must be in academic Chinese and must support later implementation, experiments, and paper drafting.

Common Mistakes

• Writing a broad review of all ideas instead of auditing one best idea
• Letting filtered or raw files override the final report
• Treating theory as decoration instead of a proof obligation
• Listing benchmarks without mapping them to claims
• Ignoring resource limits and proposing an experiment plan that cannot be run