Content
85%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 strong, highly actionable skill body: executable commands with all flags, a sequenced workflow with real validation gates (estimate-cost + explicit confirmation before spend) and error-recovery paths, and a clean two-file reference split that is clearly signaled and one level deep. The one weakness is redundancy — the multi-runtime background-download and heartbeat instructions are restated across the workflow, the command comments, and the Always Do This bullets, which inflates token cost without adding new information.
Suggestions
State the background/non-blocking download guidance (Claude Code run_in_background vs Codex foreground with yield_time_ms=1000, no &/nohup) once — either in workflow step 6 or as a single Always Do This bullet — and reference it from the Command Pattern comment instead of repeating the full text in all three places.
Merge the heartbeat bullet into workflow step 6 (or vice versa): the cadence table (under 100 -> 1-2 min, 100-1,000 -> 5 min, over 1,000 -> 15 min) and the 'report job ID, run name, output directory, next-check cadence' instruction each appear twice.
Consolidate the payload field-name guidance: 'Payload keys are molecules, target, molecule_filters — the API body field names...' in the Command Pattern section and the first two Always Do This bullets restate the same API-body-vs-CLI-flag rule three times; one statement plus the api.md pointer suffices.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense and operational with no padding explaining known concepts, but the background-download guidance is repeated three times — in workflow step 6 ('In Claude Code, use Bash with run_in_background: true. In Codex, run download-results as a foreground shell command with yield_time_ms: 1000...'), in the Command Pattern comments ('Claude Code: Bash with run_in_background=true. Codex: foreground shell command with yield_time_ms=1000'), and in two Always Do This bullets ('Prefer the agent runtime's background/non-blocking command mode for download-results... Do not append & or use nohup in Codex'), with the heartbeat cadence rules (1-2/5/15 min) and the 'report job ID, run name, output directory' instruction each also stated twice. This is real tightening work, matching 'mostly efficient but could be tightened' rather than the 'minor instances' of a 4. | 3 / 5 |
Actionability | Fully executable copy-paste commands are provided for every phase — 'boltz-api small-molecule:library-screen estimate-cost --input @yaml:///absolute/path/payload.yaml', 'start' with '--idempotency-key ... --raw-output --transform id', and 'download-results --id ... --name ... --root-dir ... --poll-interval-seconds 30' — with placeholder substitution documented in comments and concrete output paths ('-> /absolute/path/boltz-experiments/<run-name>/results/<pres_*>/...'). Specific examples cover the common cases, matching the top anchor. | 5 / 5 |
Workflow Clarity | Steps 1-7 are clearly sequenced with explicit validation checkpoints and error-recovery loops: estimate-cost with 'wait for explicit confirmation' gates the money-spending batch submit, the job ID is captured before download, auth/sandbox failures route to 'boltz-cli-setup ... before retrying', and a failed download is recovered by re-running with the same '--name' and '--root-dir' so '.boltz-run.json' resumes. The batch-operation validation cap does not apply because a cost/confirmation gate is explicitly present. | 5 / 5 |
Progressive Disclosure | SKILL.md is an overview with exactly two well-signaled, one-level-deep references, both verified to exist and to hold the bulk detail: 'Read [references/api.md](references/api.md) for the molecules, target, and molecule_filters shapes' (227-line payload reference) and 'Read [references/results.md](references/results.md) for output layout, metrics, ADME, and filtered-input accounting'. The split is appropriate — payload and output detail live in the bundle, operational guidance stays inline — and navigation is easy. | 5 / 5 |
Total | 18 / 20 Passed |