Content
78%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 a well-engineered router: domain-specific non-negotiables, a deterministic flavor-selection procedure, an action index with grep navigation, and nine verified one-level-deep reference files with explicit read conditions. Weaknesses are localized: motivational intro padding, fourfold repetition of the Scout/LGTCR pairing, a few vague action-index entries, and validation checkpoints that live in global rules and reference files rather than inline in the workflow steps.
Suggestions
State the Scout-requires-both-files rule once (in the Scout flavor bullet) and simply link back to it from Step 1, Step 2, and the reference listing instead of repeating the full pairing instruction four times.
Replace the vague action-index targets — "Challenge / remove an item → flavor reference file" and "Deploy a new registry (factory) → flavor reference file (factory section)" — with the explicit file names (light-curate.md / stake-curate.md) used by the other entries.
Add an explicit validation checkpoint to the submit/challenge/deploy workflows (e.g., a final "Validate item.json against shared-item-json.md rules and re-read live deposit before sending" step) instead of relying on the global Non-negotiables and delegated reference-file checks.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dominated by high-value, Kleros-specific operational content (the "Non-negotiables" list, flavor routing, action index) that Claude cannot be assumed to know, and the opening deposit/challenge cycle explanation is genuinely domain knowledge. However, there is measurable padding: the Scout-requires-light-curate.md pairing is repeated four times ("Three contract flavors", Step 1, Step 2, and the reference file listing: "Scout IS LGTCR at the contract layer"), and the "Why use Curate" motivational bullets ("Earn by curating", "Launch your own verification market") are user-facing rationale rather than operational instruction. This sits between 'minor instances of over-explanation' (5's neighbor, 4) and 'could be tightened' (3), clearly closer to 4 — the tightening opportunities are localized, not pervasive. | 4 / 5 |
Actionability | The body provides concrete executable guidance: copy-paste grep commands ("grep -n \"function\\|event\" references/shared-abi-fragments.md"), exact read patterns ("eth_getCode", "submissionBaseDeposit() + arbitrationCost", "arbitrationCost()"), precise structural rules ("item.json.columns must be copied verbatim from MetaEvidence", "for URL fields use type: \"link\", not url"), and the exact interactive question text to ask users. This is an instruction-only routing skill and the delegation to reference files is by design, but a few action-index entries are indirect — "Challenge / remove an item → flavor reference file" and "Deploy a new registry (factory) → flavor reference file (factory section)" name neither file nor path, unlike the entries that do. Mostly executable with minor gaps matches the level-4 anchor; it is not level 5 because the body's own executable surface stops at navigation and rules rather than covering the common transaction cases itself. | 4 / 5 |
Workflow Clarity | The three multi-step workflows are clearly sequenced with a named reference file per step ("1. shared-metaevidence.md — fetch schema... 4. shared-deposits.md — compute exact msg.value 5. Flavor reference — send the addItem transaction"), and the flavor-selection procedure has explicit fallback logic (check Scout address, pivot on PGTCR hallmarks, else default LGTCR) plus a stated one-shot correction loop. Validation is present but delegated rather than inline: the "Test the output"-style explicit checkpoint is not written into the body's workflow steps — it lives in the global Non-negotiables ("validate every metadata.columns[].type before upload", "eth_getCode before declaring any address is or isn't a contract") and inside the referenced files ("pre-upload validation", "NewItem event sampling"). Clear sequence with most checkpoints present but minor validation gaps matches level 4; it does not drop to 3 because validation steps are not missing from the bundle, only from the inline step lists. | 4 / 5 |
Progressive Disclosure | The body is a clean overview/router: it explicitly declares on-demand loading ("These 9 files are loaded on demand — only when needed... Loading an unnecessary reference file wastes context"), provides an action index mapping each operation to a specific file with grep hints, and gives a per-file description with explicit "Read this file when..." conditions. All 9 referenced files exist in references/, and their cross-references (e.g., light-curate.md → "shared-deposits.md § LGTCR specifics") are clearly signaled sibling pointers with section anchors — one level deep from SKILL.md, not nested chains. This matches the top anchor: clear overview, well-signaled one-level-deep references, easy navigation. | 5 / 5 |
Total | 17 / 20 Passed |