plan-with-me
Plan With Me
Overview
A plan is worth writing only if it removes decisions the implementer would otherwise get wrong. A plan generated in one pass from a one-line request does the opposite: it relocates the guesswork. Every unstated requirement becomes a confident-sounding sentence that nobody agreed to, and the document reads as settled precisely where it was invented.
This skill replaces that with a conversation that converges. Two rules do most of the work:
One question per turn, and every question worth the user's attention.
Every unknown is resolved by evidence or handed back to the user — never filled with an assumption.
The first keeps each round cheap enough that the user keeps engaging, and lets every answer reshape the question that follows it. Note what it does not say: that a question is only worth asking when the codebase can't answer it. Precedent in the code is a strong default, not a decision the user has delegated — where a choice is genuinely theirs, it gets asked, with the code's answer offered as the recommended option. What never gets asked is a fact you could have looked up.
The second rule is what makes the output trustworthy: a plan in which every line is traceable to a file, a source, or the user's own words.