architectlisted
Install: claude install-skill rishabhsisodiya/stop-guessing
## What this skill does
Makes one decision properly, and writes it down where it can be argued with.
An agent that builds without a decision does not stop, it guesses: it picks a library
because one was already installed, invents an enum, assumes soft deletes. None of that is
obviously wrong, which is why it survives until it is load bearing. This skill exists to
make the decision happen first, in the open.
The output is a **spec**: the decision, why, and the acceptance criteria that every later
step traces back to. `/develop` builds against it. `/verify` tests against it. `/sync`
marks it stale when the code moves on.
## What this skill refuses to do
1. **Never writes code.** Not a snippet, not a sketch, not "just the type". The moment
there is code, the decision stops being reviewable and becomes a diff.
2. **Never decides more than one thing per spec.** Two decisions in one file means neither
can be revisited on its own. Two decisions means two specs.
3. **Never presents a neutral menu.** Every choice has exactly one `(recommended)` with a
one line reason. Make the call, let the engineer override.
4. **Never leaves a value unsourced.** If an acceptance criterion needs a number, a field,
or a state, the spec names where it comes from, or records the gap as an open question.
5. **Never records a decision as `Decided` when it was assumed.** An assumption is
`Assumed` and stays flagged until it is ratified.
6. **Never edits anything outside `docs/specs/` an