convez.ai
Skill
ryanhaqueIT

prompt-coach-skill

ryanhaqueIT/prompt-coach-skill

Ver no GitHub
InicialPluginScore · de 0 a 100

Claude Code skill that coaches your prompts: fixed review contract + persistent habit log. Improves the prompter, not just the prompt.

Instalação

$npx convez-ai ryanhaqueIT/prompt-coach-skill

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
1

README

A tradução para o português deste README ainda está na fila.

prompt-coach

A Claude Code skill that reviews your draft prompt before you send it to an agent — and coaches you, not just the prompt.

Most prompt "improvers" rewrite your text and stop. This one has two jobs, always both:

  1. Improve the prompt — diagnose what would make it fail (unresolved references, unbounded scope, no deliverable, no judgment rule, missing bug symptom, stale claims, restated harness rules), then repair in the right direction. Skeletal prompts get expanded; bloated ones get elevated. Conciseness is never the goal — executability is. Prompts optimized against real evaluations usually get longer, not shorter (see GEPA, arXiv:2507.19457).
  2. Improve the prompter — every use ends with exactly one teaching note naming the habit behind the biggest flaw, and appends one line to a persistent log.md. When the same weakness shows up three times, the coach opens with the count. Your patterns become impossible to miss.

Output contract

Every invocation returns, in order: a one-line verdict (repair direction + biggest hole), the rewrite in a single copy-paste-ready block (unavoidable gaps marked [FILL IN — ...] instead of guessed), at most 2 questions (ambiguities are otherwise folded in as marked assumptions), and exactly one teaching note. The coach verifies your references with real lookups against your machine, but never performs the task itself and never auto-sends the improved prompt.

Install

Copy SKILL.md and log.md into your personal Claude Code skills directory:

# macOS / Linux
mkdir -p ~/.claude/skills/prompt-coach
cp SKILL.md log.md ~/.claude/skills/prompt-coach/
# Windows
New-Item -ItemType Directory -Force "$env:USERPROFILE\.claude\skills\prompt-coach"
Copy-Item SKILL.md,log.md "$env:USERPROFILE\.claude\skills\prompt-coach\"

Then in any Claude Code session:

/prompt-coach <paste your draft prompt>

or just say "review my prompt: ...". The log.md here ships empty — your habit log accumulates locally and stays on your machine.

Design notes

This skill was built test-first, the same way you'd build code. Baseline runs (subagents handling "improve this prompt" with no skill loaded) showed that unaided Claude already rewrites prompts well — it resolves references, bounds scope, names deliverables. What it reliably fails to do is coach: no teaching note, no memory of your recurring weaknesses, and it sometimes starts performing the task during the review. The skill therefore encodes only what the baselines proved missing: the mandatory teaching loop, the habit log, the fixed output contract, and the verify-don't-execute boundary.

Scope boundary: this coaches the task prompts you type into a coding agent. For API/system prompts with variables, use Anthropic's prompt improver; for production agent prompts that have eval suites, use eval-driven optimization (DSPy GEPA, adk optimize).

License

MIT

Ver o original no GitHub

Bom para

Relacionados