install

Copy one line. Install the two it calls.

Detects Claude Code, Cursor, Codex, Qoder and more, then writes the skill into the right folder. Install draft-skill and tune-skill too: Step 1 drafts with one and Step 3 gates with the other, and neither ships inside this repo.

npx skills add arm-pz/prompts-to-skill --skill prompts-to-skill

JavaScript is off, so the tabs above are inactive. Every install option, copied straight from the repo:

npx skills add arm-pz/prompts-to-skill --skill prompts-to-skill does it in one command: detects your agent and writes the skill into the right folder.
git clone https://github.com/arm-pz/prompts-to-skill, then copy the prompts-to-skill/ folder itself into .claude/skills/, .cursor/skills/, .agents/skills/, .qoder/skills/, or wherever your agent loads skills from. SKILL.md sits at the repository root, so there is no nested skills/ path to copy.
It needs two other skills, or it stops and says so: npx skills add arm-pz/draft-skill --skill draft-skill and npx skills add arm-pz/tune-skill --skill tune-skill.
Step 3 runs python <path-to>/tune-skill/scripts/check_skill.py <skill-folder>. That script ships inside tune-skill, not here.
Or paste to your agent: "Install the prompts-to-skill skill from github.com/arm-pz/prompts-to-skill into this project's skills directory, then install draft-skill and tune-skill too, and confirm all three load."

Browse the source: github.com/arm-pz/prompts-to-skill

the pipeline

Four steps, two confirmations, two gates.

SKILL.md is 54 lines and fixed: the order and the gates do not change to suit the skill being built. Two files in references/ hold the rest — pipeline.md for each step, gates.md for the pass/fail rules.

Before you start 4

Four checks, and nothing is drafted until the first two pass.

3 prompts minimum
screen for unsafe requests
search installed skills for overlap
confirm intent — point 1

Step 1 · Draft draft-skill

Its resource-planning and writing steps apply; its own intent and boundary steps are already settled, so it does not re-ask you.

folder = name = H1
description lists real contexts
under ~500 lines
detail into references/

Step 2 · Test cases ≥8

Two outputs, shown to you to add or remove cases — confirmation point 2.

3 clear triggers, your words
2 implicit triggers
3 near-misses
evals/evals.json

Step 3 · Tune tune-skill

Follow tune-skill's workflow on the draft, then run its validator against the folder and fix every FAIL.

collect, map, audit
edit minimally
check_skill.py
fix, re-run, record

Step 4 · Name and list optional

Only if you plan to publish. Runs findable, and a rename here means re-running the validator.

keep or rename, with the cost
About, topics, README opening
audit marked pending if no live repo

The report 7

Short, and it says what was run and what was not — including the prompts the finished skill does not cover, and why.

name, folder, repo
intent confirmed, overlap found or not
one line per step
Gate A and Gate B, with counts
prompts not covered
findable: done, pending, or skipped
one next step

ground rules

What it refuses to do.

No duplicate skills

Before drafting, it searches the skills you already have for descriptions that claim these requests. If one of them already does the job, it tells you and asks whether to tune that one instead of writing a second.

A missing dependency stops the run

Step 1 is draft-skill and Step 3 is tune-skill. If either is not installed, the skill says so and stops rather than improvising their steps. That is a real requirement, so it is stated up front, not discovered halfway.

No gate, no hand-back

It never reports a skill as complete while a gate is failing, and Gate B cannot pass on written-only cases. Unrun checks are labelled unrun — the output states what was executed and what was not.

You paste the same prompts every week.
Turn one into a skill that fires on them.