Content
90%Weight 40%Scale 1-5Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.
A dense, high-signal skill that captures genuinely non-obvious operational knowledge about Pi extension authoring with executable validation commands and a complete session-safe code pattern. The main gaps are minor: a few unsequenced imperatives in the Validation section and no explicit failure-recovery loop connecting validation output to the Common fixes table.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is lean and assumes Claude's competence: rules capture non-obvious, project-specific facts (stale ctx after session replacement, 'pi.addTool() is gone', lazy optional dependencies, Bun's minimum-release-age lag) with no concept explanations or padding. Legacy package info is properly flagged as deprecated-style notes ('older installs used @mariozechner/pi-coding-agent'), satisfying the time-sensitive-information exception. | 5 / 5 |
Actionability | Commands are copy-paste ready (bun --check, PI_OFFLINE=1 pi --help with an rg filter, pi update, npm install -g, smoke-test invocations) and the TypeScript pattern is executable apart from deliberately named user-substitution points (buildMessageWithoutCapturedPi, <extension-name>). Specific API names (ctx.newSession(), withSession, pi.registerTool(), createRequire()) and a symptom→cause→fix table cover the common cases comprehensively. | 5 / 5 |
Workflow Clarity | There is a clear ordering ('Run the cheap gates first') with explicit validation checkpoints (rg over pi --help output, which -a pi / pi --version / pi list verification, smoke tests) and a Common fixes table that serves as error recovery. It falls short of 5 because some steps ('Clean untracked npm lockfiles...', the Bun symlink guidance) appear as unsequenced imperatives and validation failures are never explicitly looped back into the fixes table. | 4 / 5 |
Progressive Disclosure | Sections are well-organized (Canonical surfaces, Rules, pattern, Validation, Common fixes) with no nested references and everything one level deep. However, at ~135 lines with no bundle files, all detail (package-update commands, the full code pattern) is inlined in SKILL.md where the rubric expects an overview pointing to detailed materials, so it is not a clear anchor-5 split. | 4 / 5 |
Total | 18 / 20 Passed |