Write your own skill

The Lean job: turn a procedure into standard work an assistant can follow

You already know how to write a good standard work sheet: the sequence, the key points, the stop rule, the record. A skill is that sheet, in a form the assistant reads when a request matches it. Treat it like standard work. It has an owner, a revision, a date it was last checked at the gemba, and a way to be improved. The difference is the reader. A person following a sheet notices when a step doesn't make sense. An assistant following one improvises, and may quietly drop a step.

Predict first

Here is an invented SOP for a first-piece check on a heat sealer. You will turn it into a skill. It is also a downloadable text file, SOP-L3-014-rev-C.txt.

SOP-L3-014 rev C (invented for training)
Line 3 first-piece check
Owner: Line 3 supervisor
Purpose: confirm the heat sealer is making good seals after a changeover.

1. Match the job number on the traveler to the schedule.
2. After any changeover, keep the first 3 pieces.
3. Measure the seal width at the left, centre and right of each piece. Spec: 12.0 mm plus or minus 0.5 mm.
4. Record the readings on the check sheet.
5. If any reading is out of spec, stop the line and call the lead.
6. The operator signs the sheet. The lead initials all sheets at the end of the shift.

History: written for the previous sealer. Rev B added step 5. Rev C changed the spec to millimetres.
Approvals: Quality ______  Operations ______  Date ______
Revision log: rev A, rev B, rev C

Before you read on, sort the SOP into three piles: what goes into the skill, what gets cut, and anything you would not encode at all. Write down one thing in each pile, and say why.

What a skill is

A skill is a folder of written instructions that the assistant reads when it needs them. One file is required, SKILL.md: a short header with a name and a description, then the steps. A skill can also carry extra files and scripts. It is not training, it is not a program, and it is not memory.

Standard work with a job aid is the right picture. The honest break in the comparison is the one from Lesson 1: a person is accountable and notices drift, and the model may silently drop a step.

It loads in three levels

1. Always present: every installed skill's name and description, about 100 tokens each. This is how the assistant knows the skill exists.
2. When the request matches: the body of SKILL.md. Anthropic recommends keeping it under 5,000 tokens and 500 lines.
3. Only when pointed to: extra files, which are read when the steps say so. Scripts run, and only their output enters the chat.

That is why many skills cost little, why the one-line description decides which skill is picked, and why a bloated body crowds out the conversation.

The description is the trigger

The assistant matches your request to the description. Say what the skill does and when to use it, in the third person, in the words your people actually use. Anthropic's own skill-creator tool says the commonest failure is a skill that doesn't trigger, and advises wording that is specific, even a little pushy. Keep it short. Anthropic's help article puts the limit at 200 characters while its developer documentation and the open specification say 1,024, so stay under 200 and you are safe either way.

It is advice, not enforcement

Anthropic's Claude Code documentation says instruction files are context, not enforced configuration, and that there is no guarantee of strict compliance. Its own skills page puts rules that must hold every time into hooks, which are automatic checks outside the model. In Lean terms this is mistake-proofing: if a step must happen, use a sign-off the model cannot skip, and don't rely on the sentence in the file. Lessons 8 and 9 build on this.

It is static, and it is a copy

• It never learns. The file doesn't change when you correct the assistant, and a new chat starts fresh. Every improvement is a deliberate edit by its owner, which is kaizen.
• Critical rules go first. In Claude Code, after a long chat is automatically summarized, only about the first 5,000 tokens of each skill are kept.
• Your copy is a copy. claude.ai, the API and Claude Code hold separate copies. A personal skill in claude.ai belongs to one user, an API skill to a workspace, and a Claude Code skill is a folder or a git repository. Decide where the master lives, and which copy is current.

What works, and what doesn't

• Write it from watched work. In a benchmark preprint, skills written by people raised the average pass rate a good deal, skills the models wrote for themselves gave no average benefit, and short focused skills beat exhaustive ones. The figures differ between versions of the paper, so I give only the direction. Start from an SOP or notes from the floor, not from a blank prompt.
• Short beats long. In a preprint that studied 307 failures caused by skills, the largest source was skills that turned checklists into mandatory heavy work, such as excessive verification (67 cases). In Lean terms that is over-processing. Another preprint found instruction-following drops as the number of instructions rises, with the best models at 68% at 500 instructions. A 15-step SOP is nowhere near that, but a skill that tries to hold everything is.
• Consistent is not accurate. A skill makes the assistant's output more uniform. It does not make the numbers right. Lesson 5's checks still apply.
• Skills decay. When the SOP changes, or the model does, or the menu names do, the skill goes stale. My reasoning, and the reason for the review date in the label below.
• Other people's skills are software. A skill carries instructions and can carry code. Anthropic's advice is to use only skills you wrote or got from Anthropic, and otherwise to audit them like any software you install. Lesson 9 covers it.

Keep, cut, and do not encode

A skill folder for the SOP above can be this small. There are no scripts, by design.

checking-line3-first-pieces/
  SKILL.md                  (short; no scripts, by design)
  references/spec-table.md  (spec limits, SOP revision, date)
  assets/check-sheet.md     (a fixed output template)

And this is the SKILL.md:

---
name: checking-line3-first-pieces
description: Prepares and reviews the Line 3 heat-sealer first-piece check sheet per SOP-L3-014 rev C. Use when asked to run, prepare or review a first-piece or post-changeover check.
metadata: {owner: Line 3 supervisor, source-sop: "SOP-L3-014 rev C", version: "0.2", verified-at-gemba: "2026-09-15 (invented)", review-by: "2026-12-15 (invented)"}
---
# First-piece check, Line 3

## Inputs (ask if missing; never guess)
Job number, and 3 pieces x 3 seal-width readings (left, centre, right). A missing reading is [NEEDS GEMBA], never a guess.

## Steps
1. Confirm the job number against the schedule.
2. Take the first 3 pieces after the changeover.
3. Compare the readings with the spec, 12.0 mm plus or minus 0.5 mm (11.5 to 12.5 mm), as in references/spec-table.md.
4. Fill in assets/check-sheet.md.
5. Any reading out of spec: write STOP and call the line lead. Offer no workaround. One out-of-spec reading is STOP even if other inputs are missing.

## Checks before answering
Every number tagged [MEASURED], [SOP] or [NEEDS GEMBA]. Nothing outside the spec table is called spec.

## Done when
All 9 readings recorded or marked missing; verdict PASS, STOP or INCOMPLETE; open items listed.

## Out of scope
Troubleshooting, changeover steps, safety rules (point to controlled documents).

Why it looks like that

• The description says what it does and when to use it, so "run the first-piece check" triggers it and "draft a SMED plan" doesn't.
• The header carries the owner, the source SOP and revision, a version, and when it was last verified at the gemba. That is the label.
• Inputs are asked for, never guessed. A missing reading is [NEEDS GEMBA], not "typically 12.0".
• The stop rule says to stop and call, and offers no workaround.
• Done when gives the assistant a verdict to reach: PASS, STOP or INCOMPLETE, with the open items listed.
• Out of scope keeps it from improvising about changeover steps or safety. Those belong to controlled documents.
• Step 6 is missing. It goes to the owner as a question: "Is initialling sheets at the end of the shift an approval, and if so, why is it after production?"

Label it, keep one master, test before you share

• Put the owner, the source SOP revision, the version, the last gemba check and a review date in the header.
• Keep the master under document control. A zip file is a copy.
• Keep a rollback copy, keep skills few and focused, and don't let the author be the only reviewer.
• Anything that must happen every time, such as a lockout, PPE or a safety stop, stays in the controlled document and on the physical check sheet. The skill drafts and a person verifies.
• Check who owns an SOP before you convert it. An SOP written at work may belong to your employer. Under Anthropic's commercial terms customers keep their inputs and own their outputs, but that doesn't settle who owns the SOP itself.

The 17 free skills in the Lean toolkit are written in this same format. Open one in a text editor and compare it with the model above.

Setting it up

In claude.ai

1. Switch on "Code execution and file creation" under Settings, Capabilities. Skills need it. It is on by default for Free, Pro and Max, and on Team and Enterprise an owner sets it.
2. Zip the skill folder with the folder at the top of the zip: checking-line3-first-pieces.zip holds checking-line3-first-pieces/SKILL.md. The folder name matches the skill's name. Use only the header fields the specification allows (name, description, license, compatibility, metadata, allowed-tools), or the upload fails.
3. Open Customize, Skills, then the plus sign, Create skill, Upload a skill, and choose the zip.
4. Switch it on. In a new chat, ask in ordinary words. The reasoning view shows the skill being read. If it isn't used, fix the description, or name the skill in your request.
5. On Team and Enterprise plans a skill can be shared, and recipients get a view-only copy. On personal plans nothing is shared: each person uploads a copy.

Anthropic's pages don't agree on which plans can upload a custom skill, so if upload isn't available to you, paste the SKILL.md text into a Project's instructions. That should work, though I could not verify the menu name, with one difference: it is always on, where an uploaded skill loads only when a request matches it.

Elsewhere

• Claude Code: a folder at ~/.claude/skills/<name>/SKILL.md for you, or .claude/skills/<name>/SKILL.md in a project, which you can commit to git to share.
• Other vendors: the SKILL.md format is an open specification, and Claude, ChatGPT and Codex, GitHub Copilot, Gemini CLI and Cursor list support for it. Packaging, plans and limits differ, so test the skill in each tool your people use. ChatGPT is reported to offer skills on its business, enterprise, healthcare and education plans (from search extracts, and I could not confirm Plus or Pro), Google says Gems convert to skills from November 2026 on personal accounts, and Microsoft's custom skills are in preview.

Your turn: from SOP to skill, then test it

About 25 minutes, using the invented SOP above.

1. Ask for a draft with no guidance (5 minutes). In a new chat, paste the SOP and ask for a SKILL.md. Note what the assistant adds, cuts and invents.
2. Redo it with the rule (8 minutes). Ask again, this time telling it to keep the sequence, the spec, the stop rule and the record, to cut the history and approvals, and not to encode step 6. Show it the folder layout and the model SKILL.md. Compare it with your first draft, and with the model.
3. Install it (3 minutes). Zip and upload it, or paste it into a Project's instructions if upload is not available to you.
4. Test it three ways (7 minutes), once with the skill and once without:

   Prep the first-piece sheet for job 4471.

   Draft a SMED plan for the Line 3 changeover.

   Job 4471. Readings: left 12.1, centre 12.0, right 12.7 mm. First piece.

   The first should trigger the skill and ask for readings. The second should not trigger it. The third should say STOP, because 12.7 is above 12.5.
5. Record (2 minutes). Did it trigger? Did it invent a number? Did it encode step 6?

You have succeeded when the skill asks for readings, says STOP at 12.7 mm, and your own draft from step 2 leaves step 6 out and lists it as an open question for the owner. Results vary from run to run, so run the three tests more than once.

Knowledge check

Your task: write one skill from one job you have watched

1. Open the printable worksheet. Pick one job you have watched this month, and sort its steps into keep, cut and do not encode.
2. Write the description and test it on three requests: one that should trigger the skill, one that shouldn't, and one with a missing input.
3. Put the owner, the source revision, the version and a review date in the header. Send your "do not encode" list to the person who owns the procedure.

Open the printable worksheet

In the toolkit

Every toolkit skill is a finished example of what this lesson asks you to write, in plain text. Read these three before you write yours. Standard Work documents a job as sequence, standard time and standard WIP. 5S Audit scores a workplace by zone and gives a prioritized action list. SMED Setup separates internal from external tasks in a changeover and plans a shorter one.

Sources

Where the facts in this lesson come from. Facts last checked October 1, 2026. Menus, plans and limits change often, and several studies are preprints, so check the vendor's page and read the research as direction. If you find a fact that is out of date or wrong, tell me.

• Anthropic: Agent Skills overview, Skill authoring best practices, Skills for enterprise, Using Agent Skills with the API
• Claude Help: What are skills?, Using skills in Claude, Creating custom skills, Provision and manage skills for your organization, Create and edit files with Claude
• Claude Code: Skills and How Claude remembers your project
• Agent Skills specification and adopters
• SkillsBench (preprint); Agent Skills Can Be Harmful (preprint); How Many Instructions Can LLMs Follow at Once? (preprint)
• Anthropic: Commercial Terms
• OpenAI Help: Skills in ChatGPT and OpenAI Academy: Skills (read through search extracts); Gemini CLI: Agent Skills; Google: Transition from Gems to skills; Microsoft: Custom skills in declarative agents