Content
81%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.
An exceptionally actionable, well-sequenced workflow skill: exact commands, explicit checkpoints, guard rails around the destructive reset, and a verification checklist. The main cost is token weight — dated internal history and repeated justificatory prose inflate the body without adding executable value.
Suggestions
Move the dated history (2026-08-17 split rationale, 2026-08-19 measurements) out of the <context> block into a project-notes or old-patterns file, keeping only the operative rule it justifies (e.g. "the audit runs inline, never via the reflect skill").
Tighten rule 2 and the memory-store paragraph in step 4: each currently restates its rationale two to three times; one sentence of justification per rule is enough.
Fold the step-4 memory-store caveat into a single rule near the scope-of-edits rule, since both say where edits may land — this removes duplicated guidance across sections.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is mostly efficient directive content, but the <context> block carries time-sensitive internal history with specific dates ("Split out of the commit workflow on 2026-08-17", "measured 2026-08-19, one run in three...") outside any old-patterns section, and rule 2 plus the memory-store paragraph re-justify themselves at length. This is more than the minor trimming of the level-4 anchor but far from the padding of level 2. | 3 / 5 |
Actionability | Guidance is fully executable for an instruction-only skill: copy-paste commands (GIT_PAGER=cat, git reset --hard HEAD~1 with two pre-checks, git rev-parse @{u}), exact commit-message templates ("agent: apply reflect-suggested improvements"), and literal one-line utterances to say aloud at each checkpoint. | 5 / 5 |
Workflow Clarity | Seven explicitly numbered steps with validation throughout: a spoken verdict checkpoint, a self-criticism pass before touching files, two guards before the destructive hard reset, an error-handling STOP rule, and a closing verification checklist. The destructive-operation validation the rubric demands is present, so no cap applies. | 5 / 5 |
Progressive Disclosure | A single-file skill with no bundle files, organized into clear tagged sections (Overview, Context, Rules, Instructions, Verification). At ~113 lines it exceeds the under-50-line exception for a 5, and the historical-context material is the one candidate for externalization — fitting the good-structure/minor-gaps anchor at 4 rather than the clear-split anchor at 5. | 4 / 5 |
Total | 17 / 20 Passed |