Content
92%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 well-architected routing skill: the body is a decision table over a verified, complete one-level-deep reference bundle, with concrete commands and genuinely non-obvious operational defaults (RPC lifecycle semantics, safety posture). The only weakness is the Source Coverage section's placement of version/date metadata outside a dedicated compatibility-style heading, which costs it conciseness points under the rubric's time-sensitive-information rule.
Suggestions
Condense the Source Coverage section: keep the pinned baseline in one or two lines under an explicitly labeled 'Version baseline' heading and drop the redundant sentence "These releases are the compatibility baseline, not interchangeable historical APIs" plus the release-specific change list, which restates what the version pins already imply.
Trim the citation section's arXiv fetch procedure to the essential rule (cite the current version via the DOI) — the conditional journal-reference and export-api details can live in a reference file if needed.
The RPC lifecycle paragraph mixes build-time defaults with protocol semantics; moving the correlation/settled-event details fully into references/rpc.md and keeping only the one-line "correlate by id, not arrival order" summary here would sharpen the overview's token budget.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is largely lean — a routing table, terse defaults with specific flags ("pi --mode rpc --no-session"), and non-obvious knowledge Claude lacks ("Split records on \n only — Node readline is not protocol-compliant") — but the "Source Coverage" section embeds time-sensitive version numbers and dates ("0.99.2 release (2026-09-30)", "pi-subagents 0.74.0") outside a deprecation/old-patterns section, which the rubric explicitly penalizes, and the citation section's arXiv fetch procedure could be trimmed. This matches the 4 anchor (efficient with minor trimmable instances) rather than 5, and is well above 3 since no space is spent on concepts Claude already knows. | 4 / 5 |
Actionability | Fully executable guidance throughout: copy-paste commands ("npm install -g --ignore-scripts @earendil-works/pi-coding-agent", "pi install npm:pi-subagents", "pi --mode json \"List files\""), named SDK entry points ("createAgentSession()", "createAgentSessionRuntime()", "ModelRuntime.create()"), and an intent-to-file routing table. It matches the 5 anchor (copy-paste ready commands covering the common cases); there are no gaps that would drop it to 4. | 5 / 5 |
Workflow Clarity | The primary workflow — "Pick the reference before answering or coding" — is an unambiguous decision table, and the Source Coverage section adds an explicit pre-flight checkpoint ("Check pi --version, pi --help, the installed package README, and TypeScript declarations before adapting examples to another release"). No destructive or batch operations are involved, so the validation cap does not apply, and the single routing action is unambiguous — matching the rubric's simple-skill exception for a 5. | 5 / 5 |
Progressive Disclosure | A clear overview with a complete intent-to-file routing table where all 35 cited reference files exist in references/ (verified against the actual bundle) and are one level deep — spot-checks show substantive content (e.g., quickstart.md's single pointer to providers.md is normal navigation, not a pass-through chain). Nothing that belongs in a reference file is inlined, matching the 5 anchor rather than 4 ('minor organization gaps'). | 5 / 5 |
Total | 19 / 20 Passed |