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-engineered operational skill: the workflow is an unambiguous, validated pipeline with executable scripts at every step, and detail is correctly split into one-level-deep bundle files that all exist. The only cost is repetition — the already-fixed/never-revert rule and the 'poison the build' framing are each restated three or four times across the Overview, steps, and decision tree.
Suggestions
State the never-revert-a-fix / already-fixed rule once (e.g. in the three hard rules) and link to references/already_fixed.md from the other mentions, instead of restating it in full in Step 0, Step 6, and the decision tree.
Drop the duplicated '90% lever' framing: it appears in the Overview prose, the Quick Start comment, and the Step 1 heading — one mention plus the reference link is enough.
Consider collapsing the Quick Start command block or the per-step commands in the Workflow section, since they duplicate each other; keeping one canonical command sequence would trim ~30 lines without losing actionability.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is high-signal throughout — no explaining what Postgres or git bisect is, commands dominate over prose — but there is noticeable repetition that could be tightened: the never-revert-a-fix / already-fixed rule is stated in full in Overview rule 3, Step 0, Step 6, and again in the decision tree, and the '90% lever' framing appears in the Overview, the Quick Start comment, and the Step 1 heading. Not 5: these repeated passages are tokens that don't add new information for a reader who read the first statement. Not 3: the repetition is deliberate safety-rule emphasis, and apart from it every section is lean and earns its place. | 4 / 5 |
Actionability | Every step gives copy-paste-ready, fully executable commands with concrete arguments (e.g. "./scripts/fuzz_sqlsmith.sh 3600 postgres", "./scripts/bisect_run.sh repro/mybug.sql REL_17_STABLE HEAD"), the Quick Start is a runnable end-to-end script, and the env-override example ("PG_PORT=55433 PG_POISON_CACHE=1 ./scripts/build_and_start_pg.sh") covers customization. Not 4: no gaps — even failure-mode outputs are specified ("SERVER IS DOWN = good", "Cache poison: guc (or macro), not none"). | 5 / 5 |
Workflow Clarity | Eight explicitly sequenced steps, each ending in a bolded "Success:" validation criterion, with feedback loops for error recovery ("Until this holds, do not theorise about root cause", "If it says none, poisoning was disabled", dirty-tree stop-and-ask gate) and an ASCII decision tree covering both crash and clean-fuzz branches. Not 4: validation checkpoints are explicit and unavoidable at every stage, including for the risky git operations. | 5 / 5 |
Progressive Disclosure | SKILL.md is a genuine overview: the workflow stays inline while details are pushed one level deep into 9 references, 12 scripts, and 1 asset — all verified to exist — with an annotated reference index that says "Load only what you need; each file is self-contained." Not 4: navigation is easy, references are clearly signaled at the exact step where they're needed, and nothing that belongs in a bundle file is inlined. | 5 / 5 |
Total | 19 / 20 Passed |