Skill Claude CodeCodex
Finalize the just-reviewed exec task — flip its roadmap checkbox and commit the whole tree (task changes + checkbox) in ONE commit, then stop. Use when the user has reviewed the current task's diff and says it's done. This is the opt-in exec deliberately withholds ("never commit or flip on your own"); running /done IS…
Skill Claude CodeCodex
Implementation stage. Build the feature in the main session/context by default; spawn a subagent only when a piece genuinely needs it (research, wide exploration, or a bounded complex sub-problem). Use when executing an approved spec. Read only the overview plus the one relevant spec, prefer Edit over Write, and keep…
Skill Claude CodeCodex
Nauval's controlled, cost-aware feature-development workflow. Use at the start of any non-trivial build, or when the user wants to plan and ship a feature deliberately instead of handing control to an autonomous multi-agent pipeline. Routes to the right stage — sketch (brainstorm) → shard (plan) → exec (implement) →…
Skill Claude CodeCodex
Review stage, human-in-the-loop. Code review is done by the engineer — the AI advises but does not gate or block. Use after implementation, before merging. Watch for the AI's additive tendency (adding new code or functions instead of removing or reusing) and prune for maintainability, so the codebase stays something a…
Skill Claude CodeCodex
Planning stage. Turn an approved concept into planning docs, but never one monolithic PRD. Produce a small overview file plus separate per-feature spec files, so that implementation reads only the spec it needs. Use when planning a build. This structure is cheaper to read, cheaper to edit, and cache-friendly compared…
Skill Claude CodeCodex
Brainstorming and second-opinion stage, used before writing any code. The user is the source of the idea; your job is to challenge it, surface alternatives and tradeoffs, and sketch what a good plan or architecture would look like. Use when refining a rough idea into a foundational concept. Do NOT jump to…
Skill Claude CodeCodex
Testing stage. Prefer writing test CASES — unit tests, or scripted e2e cases — that exercise code and commands, rather than having the agent manually pilot the UI. Avoid loops where the agent drives Playwright and eyeballs screenshots to judge correctness: its visual judgment is unreliable and screenshots are…