Content
77%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, well-architected body: a clearly sequenced multi-wave workflow with genuine validation and feedback loops, deterministic verification commands, and a real single-writer safety model for git. The two material weaknesses are the dangling evals/ references (an entire referenced directory is absent from the bundle) and codename jargon inherited from the extraction session that costs tokens without informing execution.
Suggestions
Add the evals/ directory (content-evals.json, trigger-evals.json, README.md) to the bundle, or remove the Evals section — three of the referenced paths currently dangle, which breaks progressive-disclosure navigation.
Strip or expand the session codenames ("A1 PANEL", "S7", "A8", "A9", "R3 EXTRACT") into plain language — they are opaque outside the original session and cost tokens without aiding execution.
Move the Provenance section and the dated session stats (PR #1511, 2026-05-27) out of SKILL.md or into a clearly marked background note; they are time-sensitive and not needed at runtime. Optionally merge the duplicate sibling-boundary explanations (table + trigger list + Boundary section) into one.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense and almost every line is actionable — no explaining of concepts Claude already knows — but it carries opaque internal-codename padding ("A1 PANEL + WAVE EXECUTION + S7 DETERMINISTIC TOOL BRIDGE + A8 ALIGNMENT LOOP + A9 SUPERVISED EXECUTION", "that's R3 EXTRACT in reverse") that is meaningless without external session context, plus a Provenance section with dated session history ("PR #1511, 2026-05-27") outside any old-patterns section. Not 5 because these passages and the duplicated sibling-boundary explanation (table plus trigger list plus body text) could be trimmed; not 3 because the padding is minor relative to the overall signal density. | 4 / 5 |
Actionability | Mostly executable: concrete verification commands ("`uv run apm <verb> --help` for CLI, `grep -n src/apm_cli/` for symbols, `python -c \"import ...\"`"), a real dispatch template (assets/subagent-prompt-template.md) with explicit substitution variables ("substitutes scope + working dir + branch"), and a real deterministic script (scripts/scan-cross-corpus-drift.sh) with usage notes. Falls short of 5 because the dispatch step itself is abstract ("dispatches one task per scope using the prompt template" with no example invocation) and PROBE checks are described but not given as commands. | 4 / 5 |
Workflow Clarity | The 8-step process (PROBE → RISK-TRIAGE → WAVE-PLANNER → WAVE EXECUTION → POST-PASS → ALIGNMENT LOOP → COMMIT+PUSH → PR+SUMMARY) has explicit validation checkpoints and feedback loops throughout: schema validation of every subagent return ("Validate every return against the schema; reject malformed JSON"), a bounded alignment loop ("Any residue triggers a targeted re-dispatch (max 2 redrafts) or is escalated to maintainer"), a pre-flight gate ("Verify on a working branch (not main)", ask-maintainer guard if the backfill target is missing), and a single-writer interlock for git. This batch operation fully satisfies the validation requirement. | 5 / 5 |
Progressive Disclosure | The three core bundle files referenced (assets/subagent-prompt-template.md, assets/panelist-return-schema.json, scripts/scan-cross-corpus-drift.sh) all exist and are well-signaled one level deep, but the "Evals" section points to evals/content-evals.json, evals/trigger-evals.json, and evals/README.md — none of which exist in the bundle — and there is no references/ directory, so three referenced paths dangle. Structure is otherwise good, but broken navigation plus inline content that could live in reference files (cost model, provenance, the full sibling-contract table) keeps this below 4. | 3 / 5 |
Total | 16 / 20 Passed |