/amp <raw prompt> — a Claude Code skill
that rewrites a casually written prompt into a high-quality one by injecting the
minimal set of quality-driving words for its task type, then runs it.
The words are the value. amp grounds, scopes, and adds verification sized to the task — and deliberately does nothing to trivial or conversational prompts. It's a prompt rewriter, not a process engine.
Via skills.sh:
npx skills add aimsise/ampOr, for local development, clone the repo and symlink the skill into your
Claude Code skills directory (the skill lives under skills/amp/):
git clone https://github.com/aimsise/amp ~/src/amp
ln -s ~/src/amp/skills/amp ~/.claude/skills/amp/amp add CSV export to the reports table
amp classifies the prompt (trivial | implement | debug | design | research | review), splices in the right-sized word set (see
skills/amp/references/wordbank.md), shows the enriched prompt on
one line for transparency, and runs it. Trivial and conversational prompts pass
straight through untouched.
- Classify the task in one line.
- Right-size — inject the smallest set that still grounds, scopes, and checks. Heavy orchestration on a small task lowers quality, so it's omitted.
- Splice the words into one enriched prompt (source-of-truth grounding, explicit scope, task-specific verification, stop-before-irreversible).
- Run it — surfacing and confirming first if the request is irreversible or outward-facing.
Right-sizing overrides everything. amp is opt-in, lean, and self-contained (core tools only; it never calls another skill).
In a blind head-to-head against an always-on prompt optimizer, amp won every
category (mean result-quality delta +1.38 vs +0.26, judged blind across three
lenses). The reusable harness, test set, and full results live in
evals/.
The edge is right-sizing: always-on optimizers over-structure trivial, social, and ambiguous prompts and go net-negative there; amp suppresses on those and amplifies only where it helps — and because it's opt-in, it costs nothing on the turns it shouldn't touch.