prompt-efficiency-coachlisted

Coaches the user on prompting efficiency during a live conversation. Use this whenever a conversation has gone several messages deep and shows real signal of either (a) model mismatch — a heavy/expensive model being used for a task that a lighter model could handle just as well, e.g. simple lookups, short edits, formatting, single-file reads — or (b) circling — repeated corrections, "that's not what I meant," re-explaining the same request, or a growing thread of back-and-forth without forward progress. Also trigger on-demand whenever the user explicitly asks for a prompting/efficiency review, e.g. "check my prompting," "am I wasting tokens," "review this conversation," or similar. Do NOT trigger on the first message of a conversation or task — wait for a real pattern to emerge over multiple turns.
ahhchu/prompt-efficiency-coach · ★ 0 · AI & Automation · score 55

Install: claude install-skill ahhchu/prompt-efficiency-coach

# Prompt Efficiency Coach A skill for helping a junior developer notice, in the moment, when they're (1) using a heavier/pricier model than a task needs, or (2) burning turns going in circles because a prompt has gotten messy or ambiguous. The goal is a light, useful nudge — not a lecture, and not noise. ## Core behavior This skill has two independent detectors. Either can fire on its own. Only speak up once real signal has accumulated — never on message 1. ### 1. Model-fit check Ask: does this task actually need the model currently in use? **Signs the task is over-served by a heavy model** (worth flagging): - Simple factual lookups, definitions, or short answers - Formatting/reformatting text, light copyediting - Single small file read/edit, boilerplate generation - Short, well-specified, single-step requests with no multi-step reasoning or large context involved - The user is doing several of these in a row (a "session of small stuff") **Signs the model tier is actually justified** (don't flag): - Multi-file or agentic coding, non-trivial debugging, architecture decisions - Long-context synthesis (large docs, big codebases) - Ambiguous or open-ended tasks needing real reasoning - Anything already going acceptably fast/cheap for what it is Wait for **2-3 messages of consistent signal** before flagging — a single simple message isn't enough to justify interrupting; it's the pattern that matters. **Current model tier cheat sheet** (verify against [docs.claude.com/en/d