Content
93%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 lean, highly actionable overview that correctly delegates implementation to the four bundled scripts, with a genuinely useful pitfalls section covering non-obvious contracts. Only minor coherence gaps — the "handle all three" antecedent and unexplained flags like --max — keep workflow clarity from full marks.
Suggestions
Fix the "The scripts below handle all three" antecedent — it follows four numbered steps, so say "all four steps" or "all three scripts" — and add a line on what the agent should do when a watcher exits non-zero inside a cron tick (retry, skip, or report the error).
Briefly state what "--max" does (or point readers to the scripts' --help), since it appears in the primary usage example unexplained.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Every section carries non-obvious contract information ("First run records a baseline", "max 500" ID cap, "Empty stdout on no-new", sandbox-writability caveat) and nothing explains concepts Claude already knows like what RSS or cron is. The one-line intro overlaps slightly with later sections but serves as a legitimate overview, fitting anchor 5's "every token earns its place" rather than anchor 4's trimmable over-explanation. | 5 / 5 |
Actionability | Three copy-paste-ready commands with real values ("--name hn --url https://news.ycombinator.com/rss --max 5", "--id-field event_id --items-path data.events", "--repo NousResearch/hermes-agent --scope issues"), concrete state-inspection commands ("cat", "rm"), and a ready-made cron prompt covering the summarize-or-stay-silent contract. Examples cover the common cases for all three scripts, matching anchor 5. | 5 / 5 |
Workflow Clarity | The Mental model sequences fetch → compare → write → emit and the usage-to-cron wiring is unambiguous, but "The scripts below handle all three" immediately follows four numbered steps (ambiguous antecedent — three scripts or three steps?), and beyond "Non-zero exit on fetch errors" there is no explicit error-recovery guidance (e.g. what the agent should do when a poll fails inside a cron tick). This fits anchor 4 (clear sequence, minor validation gaps) rather than anchor 5's explicit feedback loops. | 4 / 5 |
Progressive Disclosure | All four bundle files referenced in the body (watch_rss.py, watch_github.py, watch_http_json.py, _watermark.py) exist in scripts/ and are referenced one level deep via a script → use → dedup-key table and a named install path. Implementation detail stays in the bundle while SKILL.md remains an overview ("See any of the three reference scripts"), matching anchor 5; nothing inlined belongs in a separate file. | 5 / 5 |
Total | 19 / 20 Passed |