菜单
Turn verified state into execution-ready prompts with wave structure, hard rules, acceptance criteria, and honest capability labels.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | Phase Prompting |
| description | Turn verified state into execution-ready prompts with wave structure, hard rules, acceptance criteria, and honest capability labels. |
Use this skill when the task is to write or revise:
Write from current truth, not from assumptions. Start from what is already verified, distinguish what is only inferred, call out overclaimed, and identify what still blocks delivery.
Follow this structure:
Start from:
Do not write prompts that assume a summary is correct when code or live behavior says otherwise.
The actual prompt handed to an executor should usually be one round:
2-3 hours5-7 wavesIf the next step depends on whether A passes or fails, write Round 2A and Round 2B instead of one oversized prompt.
Do not write rounds that mostly:
Prefer rounds that close one real gap:
Unless explicitly narrowed, each round should try to advance:
Require prompts to distinguish:
supportedpartialbest_effortmanual-onlyblockedunsupportedNever let a prompt reward inflated claims.
For heavy operations, long context, or risky changes, prompts should include:
When producing an execution-ready prompt:
phase-XX-round-Y.md, phase-XX-round-Ya.md, phase-XX-round-Y-revision.mdFor each wave, include:
Good waves:
Weak waves:
When the round depends on a high-risk result, write the branch into the prompt:
Do not make later waves depend on a success path that may never happen.
Every prompt should require a final report with:
Modified filesWave-by-wave outcomesNew user-visible capabilitiesNew/updated APIsTests runBuild/check runRelease-readiness deltaRemaining blockers before deliveryWhy this phase materially advances toward deliveryWhen using this skill, produce one of:
Round 1 prompt plus branch promptsThe output should be ready to hand to an executor with minimal editing.
Systematically trace bugs backward through call stack to find original trigger
Deploy applications and infrastructure to Cloudflare using Workers, Pages, and related platform services. Use when the user asks to deploy, host, publish, or set up a project on Cloudflare.
Use multiple AI agents to investigate and fix independent problems concurrently
Turn verified project state into execution-ready phase prompts with wave structure, hard rules, acceptance criteria, and delivery reports. For writing prompts that drive real implementation progress, not design docs.
Receive and act on code review feedback with technical rigor, not performative agreement or blind implementation
Create isolated git worktrees with smart directory selection and safety verification