Feature Prompt

Turn rough feature requests into implementation-ready prompts.

Main session style: free-form intake first, targeted follow-ups only when needed. Accept whatever the user gives you — a sentence, a paragraph, a brain dump — parse it into the section structure, infer what you can from cwd and a code scan, and ask focused questions only for the parts that are genuinely missing or ambiguous.

Keep the final prompt terse, exact, and free of filler — but write it in plain English. The final prompt is a generated artifact and must not be caveman-compressed, even when caveman mode is active for chat.

Save the final prompt to docs/prompt/NNNN-<feature-slug>-prompt.md (see ## File Output), then ask the user to pass it to the grill-with-docs skill. A draft → review round is optional — offer it for non-trivial prompts, skip it when the inputs were already clean.

Agent Use

When the active agent runtime supports sub-agents and the user has allowed them, use read-only explorer agents where they can improve the prompt without slowing the interview.

• Use multiple explorer agents in parallel for independent codebase questions, especially multi-project mapping, integration points, existing patterns, constraints, risks, and acceptance evidence.
• Keep the main session responsible for user questions, judgment, and the final prompt.
• Do not delegate the final prompt to a sub-agent.
• Do not use worker agents or make code edits.
• Treat explorer results as supporting evidence; ask the user to confirm product intent before locking a section.