Content
86%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: copy-paste commands, concrete formats and thresholds, honest edge-case handling (empty graveyard, when to leave it buried), and correct use of a one-level-deep reference for the taxonomy. The only deductions are trimmable stylistic voice and the absence of an explicit validate-retry loop around report and plan outputs.
Suggestions
Trim the rhetorical asides (e.g. "A wall of 23 tombstones kills the funeral", "Two questions is a conversation; five is a deposition") to their underlying rules ("cap tombstones at 6-8 above ~10 corpses", "ask at most 2-3 interview questions") to cut tokens without losing guidance.
Add an explicit validate-and-retry checkpoint after producing the tombstone report or resurrection plan — e.g. re-run the scan with --json and confirm the chosen corpse's pulse/cause matches what was reported before committing to a plan.
The world-check step says "Search whether what blocked it got easier" without naming the search mechanism; specify the tool or search pattern to use so the step is as executable as the rest.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is efficient and assumes Claude's competence — no explaining of git, python, or basic concepts — with rules encoded densely (flag table, gotchas, epitaph rules). It loses a point to stylistic flourish that could be trimmed ("A wall of 23 tombstones kills the funeral", "stone is earned, not sprayed", "Two questions is a conversation; five is a deposition"), which is voice rather than instruction, matching anchor 4 rather than 5. | 4 / 5 |
Actionability | Fully executable throughout: copy-paste commands with all flags (`--days`, `--me`, `--state`, `--mark-resurrected`), a concrete ASCII tombstone format with a hard 44-column constraint, exact plan rules ("At most 7 concrete steps, ending at *shipped*", "Step 0 is always: confirm it still runs"), and example dialogue lines. Specific guidance covers the common cases end-to-end. | 5 / 5 |
Workflow Clarity | The multi-step sequence is clear (run scanner → read report → interview low-confidence verdicts → write tombstone report → world-check → resurrection plan → record with --state → relapse watch) with several checkpoints (check the evidence before repeating a cause, step-0 install verification, ask before touching the repo, ownership-filter gotcha). It falls short of anchor 5 because there is no explicit validate-and-retry feedback loop around the report/plan outputs. | 4 / 5 |
Progressive Disclosure | SKILL.md is a well-sectioned process overview; the detailed cause taxonomy is correctly split into a real one-level-deep reference (`references/causes-of-death.md`), explicitly signaled twice with when to read it ("read it before writing the report", per-cause resurrection angles), the scanner lives in `scripts/graveyard.py`, and a Files section maps the bundle. Both referenced paths exist on disk. | 5 / 5 |
Total | 18 / 20 Passed |