Transform a goal into a requirements-to-implementation workflow with verification and bounded fixes.
npx skills add https://github.com/CopeeeTang/copeee-skills --skill ultra-plan
Please help me install this Agent Skill. Skill: ultra-plan Source: https://github.com/CopeeeTang/copeee-skills/tree/main/skills/ultra-plan Command: npx skills add https://github.com/CopeeeTang/copeee-skills --skill ultra-plan
It solves the problem of turning an unclear or multi-step goal into a complete, verifiable implementation workflow without repeatedly interrupting the user—by generating a requirements spec, executing it, evaluating results with acceptance checks, and performing bounded fixes when failures occur.
Use ultra-plan when you want the assistant to clarify requirements and then automatically execute, test, review, and (if needed) fix the work until it produces a final report for your task.
High-intent search queries and answers about installing and using this skill.
To automate a spec-to-fix loop, ultra-plan transforms a goal into a requirements spec, executes it via Codex, evaluates results against acceptance criteria, and performs up to two bounded fix iterations autonomously.
An auto-executable acceptance criteria section defines specific CMD or METRIC checks that automated testing runs to determine pass or fail, ensuring the end-to-end implementation workflow produces verifiable results without manual inspection.
Yes, the autonomous workflow defaults to minimizing user involvement for multi-step development tasks, but a collaborative mode is available to prompt you for step-by-step confirmations during execution and review.
The two-phase evaluation runs automated acceptance checks first, then applies a judge review using structured findings to identify blockers or major issues, triggering a bounded auto-fix loop if failures occur.
The bounded remediation auto-fix loop is limited to a maximum of two fix iterations, meaning if blockers or major issues persist after two Codex-driven execution attempts, the process stops and reports the findings.
No, the orchestration workflow uses role-symmetric guardrails where Claude acts as the planner and judge while Codex acts as the executor and fixer, ensuring the planner never directly modifies code during review decisions.
Kafil99Generate detailed feature-level specifications in Markdown from high-level requirements.
wthrkRun self-review and Codex review in parallel and output a reconciled report.
starlink-awakenAutomate iterative optimization loops using metrics or LLM-guided evaluations.
claude-office-skillsOutput a numbered execution plan and wait for user confirmation before proceeding.
