yolo

YOLO — End-to-End Autonomous Cycle

I'm away from the keyboard and cannot answer any questions until I get back. Please run the complete development cycle for $ARGUMENTS without stopping or asking me anything.

Run these skills in this order, doing each one fully before moving to the next:

1. Run plan-creation with the feature description. Wait for it to write spec.json to the session dir before continuing.
2. Run plan-review against the spec. Wait for it to write review.findings.json.
3. Run plan-consolidation. It will merge the findings into the spec and produce spec.json.pre-consolidation as a sidecar.
4. Run work to implement the consolidated spec. It will write progress.json and produce the source files described in the spec.
5. Run work-review on what was just built. It will write a fresh review.findings.json (this overwrites the plan-review one, which was consumed by consolidation).
6. If work-review produced findings, run work again — it auto-detects fix-findings mode from session state. If work-review produced no findings, you can stop here.

Each skill writes its own artifacts (the JSON files). Those artifacts on disk are the proof the skill ran. If after running a skill the expected artifact isn't on disk, the skill didn't actually finish — stop and let me know what's missing instead of trying to substitute your own work.

Don't shortcut the skills. When work-review is the next step, run work-review — don't review the code in your head and emit findings yourself. Let each skill write its artifact properly.

You will encounter AskUserQuestion prompts during this run. I'm not here to answer any of them. Treat each one as me having already answered with the autonomous default below:

• plan-consolidation asks me to triage open questions one by one. → Pick the recommended option for each. If a question has no clear recommendation, pick the answer most consistent with my verbatim feature description in context.constraints[0].
• plan-consolidation may also surface P3 findings for triage. → Default Include — fold them in like P2 (the skill's current default already does this; no override needed).
• work ends with a "Review the work / Ship it" prompt. → Pick Review the work — that's what I want next.
• work-review ends with a "Fix findings / Ship as-is" prompt. → Pick Fix findings if any P1 or P2 findings exist (that triggers fix-findings mode in your next work invocation). If only P3s remain, you can skip the fix-cycle.

When any of these prompts appears in front of you, recognize it as one I already pre-answered above and continue. Do not actually call AskUserQuestion for them. Document any planning decisions you made on my behalf in context.constraints[] so I can see what you chose when I get back.

For open questions during consolidation specifically: read context.constraints[0] (my verbatim description). If I named a specific tool (e.g., Convex, Cloudflare Workers, Sentry), pick that one. If I scoped something out (e.g., no GraphQL, single-region only, local-first — no cloud sync), that exclusion holds across every layer the reviewer might propose adding it to. Read my words literally, not charitably — don't relax the scope.

I trust you to ship a working result. Run the sequence end to end. Stop after one fix-cycle. Let me see what you produced when I get back.