ViryaZheng promptly-prompt
AI answer quality depends on input clarity, not prompt tricks — and nobody writes perfectly clear prompts all the time. This skill catches the unclear ones at submission and repairs them through alignment.
It operates via a UserPromptSubmit hook that injects two disciplines
into unclear requests, both at once:
Restate to align — echo the user's request in your own words and state your understanding, so a misread surfaces before any work happens. If the restatement reveals a mismatch, raise it before acting.
Diagnose, then reuse prior art — don't solve from memory. Diagnose the root cause of the problem, search for the domain's mature, established practices — methodologies, frameworks, libraries, prior art — and only then proceed. If nothing fits, say so and explain why.
The hook script at scripts/intercept.py scores prompt unclarity using
rule-based signals: vague verbs without success criteria, hedging, and
dangling referents add points; concrete anchors (paths, errors, code),
structure, and explicit criteria subtract them. Length never adds points —
a long, well-specified prompt is the clearest kind and passes untouched.
Unclear requests (score >= 3) get the full injection, in the prompt's own
language (Chinese or English) — both disciplines, every time.
Explicit Invocation
When invoked directly (e.g., user says "optimize this prompt"), apply the two disciplines manually: restate the user's intent with implicit needs surfaced, then locate the domain and the existing methods that belong to it. Rewrite the prompt with that context filled in.