Plan → Goal

A rough plan and a goal spec are not two artifacts you convert between. They are two halves of one thing: the plan says how to do it, the goal says what "done" looks like. This skill bridges them — it takes a plan that is probably still coarse, fleshes it out, exposes its holes, makes the user adjudicate the ones that matter, and only then emits a goal with machine-verifiable completion criteria.

The reason it exists: handing a rough plan straight to an autonomous run means the agent spends that run guessing the user's intent mid-execution, and a wrong guess isn't discovered until tokens and file edits have already happened. Fleshing the plan out first moves every guess to the cheapest possible moment — before anything runs.

Output Language

Match the language of the user's request, and apply it to all user-facing output — option labels, generated-document headings, table column names — not just prose. If the user explicitly asks for another language, that wins.

Language follows the request, not the source material. When the user writes in Chinese but the plan, code, or reference is in English, output stays Chinese.

If the request is in Chinese, use Traditional Chinese (Taiwan usage) and keep established technical terms in English.

The English in this file is structural labelling for you, not literal output. Never mirror this file's language into your response.

The workflow

Two phases with a gate between them. Phase 1 produces a review the user reads; the model then stops and waits; only after approval does Phase 2 emit the goal. Running both in one breath throws away the entire value of the skill — the gate is where a wrong assumption costs one sentence instead of a whole run.