A workbench for developing pattern-authoring — a Claude Skill for writing other skills as Alexandrian patterns rather than procedures.
A skill written as a bare list of steps fails two ways at once. It over-applies, enforcing a rule in cases where the rule's surface conditions hold but its purpose is absent. And it under-applies, going silent on cases the author never enumerated, because the agent has a list to match against instead of a principle to reason from. Both failures are confident, and both report success.
pattern-authoring addresses this by requiring a skill to state the forces each rule resolves — the competing pressures, which one wins, and the conditions under which the rule stops binding — in six-part form: Problem, Context, Forces, Solution, Rationale, Resulting Context. It then runs five adversarial probes aimed specifically at defects that survive rereading.
See ORIGIN.md for the verbatim record of how the skill was built, including the probe pass that changed four of its rules.
The skill is a single folder: pattern-authoring/. Copy it into a repo's skills directory.
Per-project — available only in that repo. Run from the target repo's root:
mkdir -p .claude/skills && cp -r /path/to/agentic-pattern-language/pattern-authoring .claude/skills/User-level — available in every project:
mkdir -p ~/.claude/skills && cp -r /path/to/agentic-pattern-language/pattern-authoring ~/.claude/skills/Restart Claude Code, or start a new session, for the skill to be picked up.
Claude Code derives the skill's identity from its directory name, and SKILL.md declares name: pattern-authoring in its frontmatter. Rename the folder and the two disagree. Copy the folder, don't copy SKILL.md into a folder called something else.
Per-project is the safer default. A skill installed user-level is loaded in every session, including ones editing skills where you'd rather it not participate — and its description is written to fire on "writing or revising a SKILL.md," which is a real trigger you may hit while not wanting it.
There is deliberately no .claude/skills/ here. This repo is where the skill is developed, not used — and a skill under .claude/skills/ loads into every session in its repo. Installed here, it would be steering the reasoning meant to be revising it critically.
So pattern-authoring/SKILL.md is a source artifact in this repo: inert text to edit and attack. It becomes a skill only at its destination.
The skill carries five guessed forces — inferences stated decisively in the text but never confirmed against practice. They are listed at the end of ORIGIN.md and remain open. Two of them, the cost asymmetry that justifies the whole approach and the claimed frequency of decorative forces, are load-bearing: if either is wrong, the method is heavier than the problem it solves.
It also carries one unpatched limitation. The skill instructs an author not to invent a rationale for a rule they don't understand, and to record it as a guess instead. But a document that fabricates one is indistinguishable from an honest one. The guessed-forces list is the only check, and it is honesty-dependent with no mechanism behind it. This is recorded rather than patched, because a rule added to sound reassuring would be exactly the defect the first probe exists to catch.