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 well-structured operating guide: the Step 0–7 flow is explicit with human-in-the-loop checkpoints and feedback loops, and bundle layout is exemplary — every reference is purpose-annotated, real, and one level deep. The main weaknesses are repetition (core rules restated three to four times across philosophy, steps, contract, and nevers) and motivational framing that could be tightened, plus landing commands deferred to a reference rather than inlined.
Suggestions
State the never-write-main / never-auto-merge rule once in 'Hard nevers' and cross-reference it from S5, Step 6, and the invocation contract instead of restating it in each place.
Compress the 'The bigger picture — say it to the human, because it's the point' paragraph to a single sentence and cut the repeated 'clean trail produces nothing' framing (keep it in S4 or Hard nevers, not both plus Step 5).
Inline the exact detached-checkout + push command sequence for the default landing path in Step 6 (currently only in references/outcomes.md) so the most common path is copy-paste ready from the body.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is mostly project-specific and assumes Claude's intelligence (no generic concept explanations), but the same rules are restated repeatedly — 'never write main / never auto-merge' appears in S5, Step 6 ('Never write `main`; never auto-merge'), the invocation contract ('never on `main`'), and Hard nevers; 'a clean trail produces nothing' appears in S4, Step 5, and Hard nevers — plus the motivational 'The bigger picture — say it to the human, because it's the point' paragraph. Not a 2: the length is largely load-bearing discipline for this project, not padded teaching of known concepts. | 3 / 5 |
Actionability | Concrete, runnable guidance throughout: 'scripts/resolve-sessions.sh <PR#|feature>', triage by 'high `Attempt`, non-zero `Exit`', four named lenses, 'evals/<name>/run-eval.sh', 'git push origin HEAD:<branch>', 'branch `capture/<slug>` off `origin/main`', 'git checkout main'. Not a 5: the exact detached-checkout command sequence and eval authoring detail are deferred to references, so the most common path is not fully copy-paste ready from the body alone; not a 3: the mechanical steps that are inlined are executable as written. | 4 / 5 |
Workflow Clarity | Steps 0–7 are explicitly sequenced with validation checkpoints at every risky juncture: Step 0 clean-tree preflight, Step 2 'Tell the human your triage in two lines before diving', Step 4 joint defect-vs-inherent-difficulty classification, Step 6 'Confirm the human's explicit choice' before any landing, and Step 7 return to main. Feedback loops are present (missing-trace degradation note, right-reason eval check, classification loop) and 'Hard nevers' serves as a checklist. The destructive/batch cap does not apply: git operations are gated by the preflight check and explicit human confirmation. | 5 / 5 |
Progressive Disclosure | A clear overview body points to four well-signaled references, each annotated with its purpose and a 'hackable seam' note (trace-reading, evals, outcomes, observability-tooling) plus one script; all cited paths exist on disk and none of the reference files cross-reference each other, so references are one level deep. Detail is appropriately deferred ('Full detail and the project-folder layout live in `references/outcomes.md`'). Not a 4: navigation is easy and the split is clean, with no orphaned or buried content. | 5 / 5 |
Total | 17 / 20 Passed |