Run · Chapter 02

Skills and playbooks

Turn demonstrated work into reusable, reviewed agent procedures.

A well-done job usually contains choices that its instructions leave out. Someone notices an exception, asks a clarifying question or knows when to stop. A reusable skill makes those choices available for another agent to follow and for a person to inspect.

Begin with one recent example of the work. Ask the person who did it to narrate the trigger, inputs, actions, decisions and result. Pause where they say “usually” or “obviously”: that is often where an important rule is still tacit.

Turn the demonstration into a procedure

Write a focused SKILL.md describing when to use the method, the information it needs, the steps it follows and the quality checks it must pass. Include tool requirements, exceptions and the point at which a person decides. The Agent Skills specification gives the file a portable structure; actual discovery and installation still depend on the assistant.

Look for a fitting existing method before writing a new one. Matt Pocock’s engineering skills and Corey Haines’s marketing skills offer examples of packaging work this way. Inspect the instructions, scripts, dependencies and license before installing a selected skill. Their existence does not establish that they work for your case.

Give reuse a real test

Have another agent follow the procedure on a fresh permitted input. Observe where it needs help and incorporate the correction. Keep the skill’s version with the resulting evidence so a later change can be compared against it.

You have made tacit knowledge reusable when the next run produces work someone accepts and the reasons for its boundaries remain visible. If you publish the skill, prepare a separate edition that removes private material and respects the rights in the original work.

Tools and setup

Practitioners and tools: Matt Pocock's skills package engineering interviews and workflows; Corey Haines's Marketing Skills cover marketing jobs. The Agent Skills specification defines SKILL.md with name/description frontmatter. These are reusable implementations, not proof of their effectiveness for your case.

  1. Inspect the relevant skill files, scripts, source and license before installation. Pick one job, not an entire collection by default. Record the source revision; do not execute embedded instructions during review.
  2. In a local Codex project, place a reviewed skill in .agents/skills/<skill-name>/SKILL.md, preserving required support files. This repository path is documented in OpenAI's skill guide. Other assistants have different discovery paths; check their docs rather than assume portability means identical installation.
  3. Corey Haines's documented discovery command is npx skills add coreyhaines31/marketingskills --list; after review, install only the selected skills using the supported agent target. Matt's collection asks users to run its setup skill per repository. Start a fresh selection context and verify the intended skill is discovered.
  4. Narrate a real job, draft a narrow procedure, then rehearse it on a fresh input. Produce a separate redacted edition if publishing; verify rights and approval first.

Setup paths checked against linked documentation on 15 November 2026. Verify current account support during setup; reader integrations have not been tested here.

Work through this with your agent

Paste the request into your assistant. It will interview you and help produce the working artifact.

Interview me to turn ONE real job into a reusable skill. Ask at most three questions per turn and wait. First inspect authorized existing skills; reuse a fitting one. If considering third-party skills, inspect their instructions, scripts, dependencies and requested access before execution. Ask me to narrate a recent run or provide an authorized recording. Extract triggers, inputs, ordered actions, decisions, tool/account requirements, exceptions, approval boundaries and observable quality checks. Ask about tacit choices and common mistakes. Draft a focused SKILL.md with name/description frontmatter, instructions and one worked example in an agreed folder; do not overwrite existing skills. Rehearse with a fresh permitted case and record failures and corrections. Treat unexercised paths as untested. If I want a public edition, remove private data and verify rights; show it for approval before publishing. Completion means a usable skill plus rehearsal evidence, not a transcription alone.

Use these documented setup paths as references; check current account support before acting:
https://github.com/mattpocock/skills
https://github.com/coreyhaines31/marketingskills
https://agentskills.io/specification
https://learn.chatgpt.com/docs/build-skills

The Hypothesis Diff · weekly

Know when a hypothesis is stressed or changes.

One email a week: which hypotheses hold, which are stressed, which changed, and the tools worth trying. Nobody else publishes what stopped being true.