Content
67%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: triage-first structure, crisp decision tree, and consistently signaled one-level-deep references per subdomain. Its main cost is token weight — the Example Invocations section, Critical Patterns, and Anti-Rationalization table restate routing already covered elsewhere, and OS-version specifics are scattered rather than isolated in a deprecation section.
Suggestions
Cut or collapse the Example Invocations section into the Decision Tree: roughly 15 of its 17 entries repeat a decision-tree branch verbatim, removing ~50 lines with no routing information lost.
Delete the Critical Patterns section — it restates the Foundation Models Work and Diagnostics bullet lists from earlier in the same file.
Consolidate time-sensitive version facts (26-cycle adapter obsoletion, Opus 4.6 → 4.7 migration, FB18924722) into a single 'Version-boundary notes' or deprecation section so they stop penalizing the routing body's conciseness.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The routing tables are dense and useful, but the ~50-line Example Invocations section substantially duplicates the 13-step Decision Tree, Critical Patterns repeats the earlier Foundation Models Work listing, and time-sensitive version pins ("26-cycle", "27.0", "Opus 4.6 → 4.7", "FB18924722") are scattered rather than confined to a deprecated/old-patterns section. This fits the mostly-efficient-but-needs-tightening anchor rather than the minor-trimming level 4. | 3 / 5 |
Actionability | Every scenario ends in a concrete destination — named files ("Read: skills/foundation-models.md"), agents ("Launch foundation-models-auditor agent"), or commands ("/axiom:audit foundation-models") — satisfying the instruction-only guidance standard. Not a 5 because the routed-to files are not part of this bundle and several entries are prose-heavy conditional paragraphs rather than crisp directives. | 4 / 5 |
Workflow Clarity | The triage-first flow ("First, determine which kind of AI the developer needs" → triage table → training-path boundaries → numbered Decision Tree) is clearly sequenced and unambiguous. It stops short of level 5 because there are no checkpoints or recovery guidance (e.g., handling a route target that does not exist), though no destructive or batch operations exist to trigger the validation cap. | 4 / 5 |
Progressive Disclosure | References are consistently one level deep and well signaled, with per-file content previews organized by topic (foundation-models, -ref, -diag, -evaluations, -adapters) — a genuinely good hub design. Scored against the body alone since no bundle files ship with this skill; it misses level 5 because substantial inline material (the Anti-Rationalization table, External Resources, duplicated sections) is weight that belongs in the referenced files. | 4 / 5 |
Total | 15 / 20 Passed |