project

/project — the project entry point

/project is the single entry point for project work. It composes the loop's three verbs (think → do → wrap) into one skill that picks the right action based on what you asked and what stage the project is in.

The index render lives here rather than in a skill of its own because a cross-skill dispatch boundary is where the agent loses the thread: orchestration overhead causes the second skill's instructions to be skipped. Folding eliminates the boundary; the index-render is step-0.5-render-index here.

Critical

• Goal-alignment ALWAYS comes first when the think flow runs. The think flow's think-step-0-goal-alignment produces (goal in plain language + simplest mechanism + explicit ask "does this match?") and MUST wait for explicit user approval before context-load, design, or SPEC writing. Silence is NOT approval. Skipping this is the single most expensive mistake the loop makes: a goal or shape mismatch caught here costs one turn, and caught after the build it costs the rebuild. This step is not the whole enforcement: it only reaches sessions that enter /project, /spec or /implement, and most task-shaped work does not. No hook covers the sessions it does not reach, so outside this step the gate holds by judgment alone.
• SPEC writing + machine review live in /spec, not here. Per think-step-4-dispatch-spec, /project does the thinking (goal-alignment → context-load → explore → design → row split → categorize) then hands the design to /spec, which writes one spec folder per designed roadmap row, dispatches /spec-audit (reviewer consensus + spirit-check), folds findings, and auto-closes. /project never writes the SPEC itself and never runs the machine review. The user is NEVER the first reviewer — the SPEC is machine-audited before it lands, because a plan presented raw ships the findings a reviewer would have caught in one pass. Each artifact has exactly ONE machine reviewer: SPECs get /spec-audit (inside /spec); shipped diffs get /implement-audit (inside /implement). One artifact type means one machine reviewer per artifact-write.
• No copy-paste-command outputs as the default. When the router lands on ready-to-implement in a short session, INVOKE /implement [slug] inline — do not tell the user to paste a command. The only acceptable copy-paste output is the long-session fresh-tab handoff (see step-2-turn-count-gate), and only because /implement's fresh-context boundary is load-bearing: a session that wrote the plan defends it mid-build instead of judging it.
• Stage detection is a script result, not a judgment. Run scripts/detect-stage.ts to get the stage; don't infer from prose.
• Credentials in plans must resolve to a named secret-store item + field BEFORE plan-review. "From the key or whatever app Y uses" is a AGENTS.md § Standards → Read before asserting violation. Resolve the item's name + field (and its id, where the store has one) during think-step-2-explore by reading the store or the repo's credential helper, and cite it explicitly in the plan. A plan naming a credential nobody looked up names one that may not exist.
• Standards made incorrect by the SPEC's work are in-scope for the same commit. If shipping the SPEC's work makes any statement in an AGENTS.md — the root one or a folder's — (or any other standing rule the repo keeps) wrong, it MUST be amended in the same commit. Enumerate "standards made incorrect" as part of think-step-3-design. Deferring the amendment to a follow-up iteration is the deferral pattern AGENTS.md § Standards → No deferral forbids.
• Blank-mode: paste the generator output INTO your response. After step-0.5-render-index runs Bash, the captured stdout is in your context but NOT yet visible to the user. Many harnesses render the Bash tool card collapsed — a one-line chip with a disclosure arrow, no output body. So stdout left in the tool card has reached nobody; "ran the generator" is not the deliverable, the user reading the index in your reply IS. Paste it verbatim: buildCompact emits finished markdown, so there is no formatting judgment left to make — reformatting it is how the paste turns into a summary. Self-check before sending: does the reply text literally contain the project rows? The failure this prevents is the agent running the generator and then asking "which one to start?" of a user who cannot see the list.