Content
67%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 a well-structured, sequenced workflow with a concrete output contract and correctly offloads SVA syntax and the pattern library to two real, clearly linked reference files. Its main weaknesses are redundancy — the output specification appears twice and pattern listings overlap the reference file — and a worked example that stops short of showing the final generated SVA, which keeps it from being fully tight and copy-paste ready.
Suggestions
Collapse the duplicated output specification: Step 6 and the 'Output Format' section describe the same deliverables twice — keep only the format template and have Step 6 point to it.
Trim the pattern enumerations and 'Look for:' bullets in Steps 2–3 that restate domain knowledge or duplicate common_patterns.md, replacing them with a short checklist plus a link to the reference file.
Extend the FIFO example to show the actual generated SVA for one property (e.g., the assert property for 'No write when full') so the end-to-end output is concrete rather than described in prose.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Steps 2–3 enumerate pattern/property pairs with 'Look for:' bullets (e.g., 'Look for: Multiple signals that should never be active simultaneously') that restate domain knowledge Claude already has and largely duplicate references/common_patterns.md, and Step 6 plus the Output Format section specify the same deliverables twice. It matches 'mostly efficient but includes some unnecessary explanation or could be tightened' rather than level 2 because it never explains basic concepts (SVA syntax is correctly offloaded) and the conservative-inference and vacuity guidance genuinely earns its tokens. | 3 / 5 |
Actionability | Provides a concrete SVA template ('assert property (...) else $error(...)'), an explicit output format skeleton, and a worked FIFO example, which is mostly executable guidance with minor gaps. Not level 5 because the FIFO example stops at prose ('No write when full', 'Count within bounds [0:DEPTH]') without showing the final generated assertion, and not level 3 because the templates and referenced pattern library are real SystemVerilog rather than pseudocode. | 4 / 5 |
Workflow Clarity | The six steps (parse → invariants → liveness → map to properties → classify → generate) are clearly sequenced, with quality checkpoints present via guidelines ('Check vacuity: Suggest cover properties for antecedents', 'Be conservative', 'Mark confidence', and the bounded-liveness note). Not level 5 because there is no explicit validate-and-recover loop for the generated assertions (e.g., checking they compile or reclassifying a failing property), and not level 3 because those checkpoints, while not framed as feedback loops, are explicitly stated. | 4 / 5 |
Progressive Disclosure | Both bundle files (references/common_patterns.md, references/sva_syntax.md) exist, are one level deep, and are well signaled — linked inline at the point of use in Step 4 and again in a dedicated References section with descriptions. Not level 5 because Steps 2–3 inline pattern/property detail that also lives in common_patterns.md, a minor organization gap that leaves some content duplicated rather than cleanly split. | 4 / 5 |
Total | 15 / 20 Passed |