molyanov-ai-dev: Skill for Codex

User Spec Planning

Thorough adaptive interview → codebase research → completeness review → user-spec.md → three-lane validation → user approval. Output: work/{feature}/user-spec.md with status approved.

Interview Style

Conduct the interview in the language the user writes in. Be an engaged co-thinker: propose solutions, challenge weak answers with concrete examples or code evidence, and keep interviewing until the applicable requirements are understood.

• Ask 3–4 questions per batch and run as many batches as needed.
• Save every question and answer verbatim after each user response.
• Give one substantive challenge to a weak or unclear answer, then accept a supported answer and move to the next gap.
• When the user does not know, offer concrete options or break a required question down. An optional detail may remain an acknowledged limitation; a required detail may not remain TBD.
• Record material choices, rejected alternatives, and reasons in the relevant topic summary so they survive into Accepted Decisions.

When Project Knowledge exists, read its SKILL.md as the router and load only the references relevant to the task. Missing Project Knowledge does not block planning.

Workflow

1. Start or Resume

If the user explicitly asks to continue an existing user spec and provides its feature folder or slug:

1. Use that exact work/{feature} directory. Do not search for other interviews.

2. Read logs/userspec/interview.yml and the existing feature artifacts. Treat any additions or changes in the current request as interview input.

3. Derive the next action from the interview and artifacts:

   • continue with the earliest required topic below 85% or with an unresolved gap;
   • once the general task is understood, create code-research.md if it does not exist, then use it for the remaining questions;
   • when all required topics are complete and no substantive draft exists, run the completeness review;
   • when a filled draft already exists, validate it again with fresh reviewers rather than trying to restore old reviewer responses.