Goal

• Given a file <FILE> containing research ideas in a very early, unstructured form (loose fragments, one-liners, mixed bullets), add a ## <Title> header above each distinct idea, then, once the user approves, create one template-formatted research idea file per idea

Workflow

Read the Inputs

• Read <FILE>
• Read the template .claude/templates/research_idea.template.md
• Read .claude/skills/research_idea.rules.md for file naming and directory conventions
• Read the file names under research/ideas/*.md to know which ideas already exist

Title and Group the Fragments

• Split <FILE> into distinct ideas
  • Merge fragments that clearly describe the same idea into one group; it is fine to reorder or combine fragments that belong together
  • Keep fragments that are already distinct ideas separate, even if short
• Add a ## <Title> header directly above each idea's raw content
  • <Title> is a short noun phrase (3-8 words) that names the idea, in the same style as existing file names, e.g. Hierarchical Training, Compression as Proxy for Understanding
  • If a fragment already has a ## header, keep or refine that title instead of adding a duplicate one
• Do not summarize, expand, or fabricate content; only add headers and move fragments next to the other fragments of the same idea
• Flag (do not silently drop) any fragment whose idea already exists as a file under research/ideas/*.md, so the user can decide whether to skip it
• Write the titled, grouped result back into <FILE>

Wait for the User

• Show the user the list of proposed titles and which fragments were grouped under each
• Wait for the user to review and approve <FILE> before creating any new idea file
• If the user requests different titles or groupings, apply the changes and present the file again

Create One Idea File per Approved Idea

• For each approved ## <Title> section, create research/ideas/draft.<Idea_Name>.md, per .claude/skills/research_idea.rules.md ## Status Prefixes
  • <Idea_Name> is <Title> in Title_Case_With_Underscores, matching existing file names (e.g. Hierarchical Training -> Hierarchical_Training)
• Seed the new file with the section's raw content as the seed for Core Idea
• Skip creating a file for any idea the user chose to skip in Step 3

Clean Up the Source File

• Ask the user whether to remove, from <FILE>, the sections that were split out into new idea files (or delete <FILE> if every section was split out)
• Only remove a section once its new idea file exists and the user confirms its content was fully captured

Conventions

• Follow .claude/skills/research_idea.rules.md

Constraints

• Do not create any new idea file before the user approves the titled, grouped <FILE> from the Title and Group the Fragments step
• Do not fabricate content when titling or grouping fragments; only add headers and reorganize the existing raw text
• Do not remove content from <FILE> without user confirmation, per the Clean Up the Source File step

Examples

• research/ideas/new_ideas.md: a raw idea dump file, the typical <FILE> input for this skill
• See .claude/skills/research_idea.rules.md # Examples for a fully formatted research idea

Verification

• Every fragment in <FILE> is grouped under a ## <Title> header
• The user approved the titled, grouped <FILE> before any new idea file was created
• Each new idea file follows .claude/templates/research_idea.template.md
• Each new idea file is named research/ideas/draft.<Idea_Name>.md per .claude/skills/research_idea.rules.md ## Status Prefixes
• No content from <FILE> was fabricated or lost while splitting