Content
88%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 is an exceptionally actionable, well-sequenced orchestration procedure with explicit validation and recovery at every stage, and its bundle references are real, clearly signaled, and one level deep. Its weaknesses are mild: a few overloaded mega-paragraphs that could live in reference files, keeping the entrypoint from being a true overview.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body assumes Claude's competence throughout — no basic concepts are explained, and nearly every token carries operational contract ('exit 75, which means retry later, not failure'; 'Corrections consume the task's allowance, failed ones included'). Not 5: the mega-sentences in steps 6, 8 and 9 cram edge cases and escalation rules into single paragraphs that could be tightened or offloaded to the references, a mild form of over-explanation. | 4 / 5 |
Actionability | Fully executable guidance: exact command templates with every flag ('factory.py admit --task TASK --line LINE --goal GOAL --repo /absolute/repo --acceptance REFERENCE'), precise failure semantics (exit 75, 'at zero, launch refuses further corrections'), bounded escalation rules ('three per task'), and concrete fallbacks ('reset a writer's checkout, then launch the replacement under a new ID'). | 5 / 5 |
Workflow Clarity | Steps are explicitly sequenced (numbered 1–9 under a router that says 'pick the step that matches the user's intent and execute only that step') with validation checkpoints and feedback loops everywhere: verify the writer's commit, route findings to correction jobs, re-verify with --from-writer, release judgment, and a final acceptance judgment. Error recovery is explicit for every failure mode (dirty worktree reset, brief-known-wrong cancel, non-convergent task escalation to architect then judge). | 5 / 5 |
Progressive Disclosure | All four referenced files exist and are one level deep with clear purpose statements and markdown links — [the command reference](references/commands.md), [the role catalog](references/roles.json), [the ready-plan reference](references/ready-plan.md), [the delivery reference](references/delivery.md) — and command I/O contracts, role definitions, plan format and delivery process are appropriately split into them. Not 5: the body itself is not an overview; it inlines dense procedural manuals (the correction/allowance/judge escalation machinery in step 8, rebase-conflict resolution in step 9) whose level of detail matches what the references already carry, rather than summarizing and linking. | 4 / 5 |
Total | 18 / 20 Passed |