Content
73%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.
The body delivers excellent actionability and workflow clarity — executable commands, phase gates, verification greps, and checklists — but token efficiency is hurt by the two headline rules being fully restated in five places each, and progressive disclosure is broken in practice because nearly all referenced rule and template files are absent from the bundle.
Suggestions
Add the missing bundle files: 8 `rules/*.md` and 5 `templates/*.md` files are referenced throughout (phase table, Required Reading, Definition of Done) but do not exist in the bundle — the scaffold workflow cannot follow phases 1–4 without them.
State the log-visibility and feedback rules once each (in their rule files or one summary section) and link to them, instead of restating both across the Non-negotiable section, phase table, Core Principles, Anti-patterns, and Definition of Done.
Trim the long Definition of Done feedback checklist item, which re-inlines the full trigger table and per-trigger reaction protocol that `rules/feedback.md` is meant to hold, down to a one-line check referencing that file.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is mostly efficient — tables, one-liners, and checklists, with no re-explanation of concepts Claude already knows — but the two headline rules (log visibility and comment feedback) are each restated across five sections (Non-negotiable, phase table, Core Principles, Anti-patterns, Definition of Done), including a ~14-line DoD item that re-inlines the trigger table and reaction protocol that rules/feedback.md is supposed to hold. That duplication goes beyond the minor trimming of the 4 anchor, but the rest of the body is tight, so it does not fall to 2. | 3 / 5 |
Actionability | Guidance is fully executable: a copy-paste `run:` block with `set -euo pipefail` + `tee`, exact `gh run list ... --jq` metric commands for review mode, an exact review report format, and a concrete list of forbidden strings (`> /dev/null`, `--silent`, `|| true` without echoing). This matches the 5 anchor — copy-paste ready commands covering the common cases. | 5 / 5 |
Workflow Clarity | Scaffold mode is a five-phase sequence where each phase has an explicit gate ("do not proceed until it passes"), batched Phase-0 questions with a "repeat the answers back" checkpoint, and a Phase-5 self-check against a Definition of Done checklist; review mode has numbered steps with PASS/WARN/FAIL line-evidence requirements and a verification grep. This matches the 5 anchor: explicit validation steps, feedback loops, and checklists. | 5 / 5 |
Progressive Disclosure | The design intent is a thin index with well-signaled one-level-deep references (per-phase rule files, per-file template descriptions), but scored against the actual bundle, 12 of the 13 referenced files are missing — the rules/ and templates/ directories do not exist, so the phases 1–4 instruction to "Walk each phase using the linked rule file" fails in practice. Only references/decision-tree.md is real, which leaves navigation broken for the bulk of the content — above the 1 anchor (the body still has real standalone structure) but within the 2 anchor's minimal effective structure. | 2 / 5 |
Total | 15 / 20 Passed |