Content
77%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-engineered router: mandatory intent classification, explicit disambiguation, gated credentials, per-track phase tables with hard verification gates, and a genuinely useful probe script. The main defects are duplicated design-match detail that belongs solely in its reference file, and — more seriously — that the bundle omits nearly all top-level execution modules the body routes to, breaking progressive disclosure for every hands-on track.
Suggestions
Ship the top-level execution modules the body routes to (RULES.md, builder.md, sdk.md, credentials.md, migrate.md, sendbird-migration.md, ai-integration.md) or trim the body to only reference files that exist in the bundle.
Move the 'Front-loaded traps' detail out of SKILL.md and leave only a 2-3 line flag pointing to references/design-matching.md, where the body already says the full detail lives — this would remove ~15 lines of duplicated context loaded on every invocation.
Consolidate repeated gate rules ('run gate commands from an absolute cd and never piped', verification gate wording) into RULES.md or a single section referenced from the M4/S5 phases instead of restating them per track.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The router body is mostly dense, actionable pointer content (classification tables, module map, phase tables) with little explanation of concepts Claude already knows. It is not a 5 because the 'Front-loaded traps' section (~10 lines of fine-grained design-match detail: 'messageComposer.wrapper', 'toggleAttachmentPicker', keyboard-offset rules) duplicates content the body itself says lives in references/design-matching.md, and a few rules are stated twice ('Run gate commands from an absolute cd and never piped' appears in both M4 and S5). | 4 / 5 |
Actionability | The body provides copy-paste-ready artifacts: a complete read-only probe script, exact status-line outputs to produce, verbatim disambiguation questions to ask, exact version pins ('bump to >=4.5.1/>=0.10.2'), and named packages per product. It is not a 5 because several load-bearing steps are pointer-only and those pointers dead-end — the central execution modules the whole flow depends on (builder.md, sdk.md, credentials.md, migrate.md, sendbird-migration.md, RULES.md) are not present in the bundle, so a user following the file literally cannot execute Tracks A/B/D/M/S end-to-end from what ships here. | 4 / 5 |
Workflow Clarity | Sequencing is explicit and gated: mandatory Step 0 classifier before any tool call, product classifier, credentials step gated before connection code, and per-track phase tables with hard verification gates (M4: 'tsc --noEmit -> Metro bundle -> native build -> simulator/device smoke', S5: ordered gates including 'zero @sendbird imports'). Failure paths are explicit ('stop and ask the user - never build... from memory') and the Sendbird-removal migration includes validation checkpoints, so it clearly matches anchor 5 rather than anchor 4's 'minor validation gaps'. | 5 / 5 |
Progressive Disclosure | The router structure itself is well designed — a module map, one-level-deep reference links, and clearly signaled per-track files — and all references/*.md links resolve. But scored against the actual bundle, 7 of the most-referenced module files (RULES.md, builder.md, sdk.md, credentials.md, migrate.md, sendbird-migration.md, ai-integration.md) plus the cross-skill ../stream/sendbird-data-migration.md do not exist in the shipped bundle, so the primary navigation paths for every hands-on track dead-end. That is more than the 'minor organization gaps' of anchor 4, but better than anchor 2 since the skill is not monolithic and the references that do exist are clearly signaled and correctly placed. | 3 / 5 |
Total | 16 / 20 Passed |