alex-review
Skill61 starsChanged 6 months agoApache-2.0View on GitHub
Review Ragu changes using Alex Xiong's priorities: adversarial arguments, invariant ownership, real costs, and concrete APIs. Use when Alex's perspective is requested.
In this repository15 other files it gives its agents.
- book-fixmeSkill
- book-refineSkill
- book-reviewSkill
- clean-circuitsSkill
- code-refineSkill
- code-reviewSkill
- doc-auditSkill
- fv-docsSkill
- fv-preflightSkill
- fv-reviewSkill
- help-book-reviewSkill
- master-fvSkill
- sean-reviewSkill
- spec-fidelity-auditSkill
- tal-reviewSkill
The file
We split this skill into the rules it gives an agent
5 rules. About 0 are common, 5 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈5
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 19
Check algebra and implementation independently.
- Distinctive: line 49
Check the consumer's required equivalence against zero constraints, nested invocations, relative wire positions, and configurable routine …
- Distinctive: line 57
Check whether benchmarks include setup and allocation, measure their named operation, and reach the claimed hot path.
- Distinctive: line 71
Read signatures as a caller: prefer clear names, flatter arguments, and existing implementations.
- Distinctive: line 88
When uncertain, briefly restate your reading and ask the specific question that would confirm or refute it.
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Alex review
Apply the relevant priorities below to the requested scope and exact revisions, using current protocol descriptions, code, callers, tests, and repository policies. Give your own review, not Alex's approval. Keep feedback local unless publication is authorized.
Reconstruct the protocol before judging the patch
Identify the objects, their mathematical meaning, and when the prover fixes values and learns challenges. Ask what the prover could choose after learning each challenge. Track commitments and evaluations across stages, accumulator updates, and both curves; confirm new state replaces old state. Check algebra and implementation independently. A convention can differ from the book without breaking correctness. Distinguish public data, binding, hiding, and non-identity requirements before removing blinding.
Test the adversarial argument, including its counterargument
Try a concrete deviation: substitute a circuit, repeat a step, move witness support, vary dimensions, or reuse a routine identity. Name the verification equation it passes and the intended equation it violates. Then seek the binding that defeats the attack. Configuration may be encoded in a registered circuit without being a public input; changing honest synthesis alone does not show verifier acceptance. Withdraw concerns defeated by that binding.
Separate honest-prover completeness from malicious-prover soundness. A local counterexample can expose an obligation without demonstrating a forged proof. Proposed fixes must use information the verifier has: commitments and openings do not necessarily reveal coefficients for a support check.
Make non-local invariants explicit
Find dependencies split across helpers, flags, callbacks, scope setup, caches, and wire remapping; illustrate them concretely. Localize transitions and their invariants, or add a targeted test. Verify shortened cache keys over the queried domain and distinguish local counts from absolute indices. Defensive fields and assertions need specific failures not already excluded by an invariant.
Define equivalence at the level that uses it
For fingerprints and memoization, distinguish equal types, input/output types, constraint counts, polynomial contributions, and interchangeable output wires. Check the consumer's required equivalence against zero constraints, nested invocations, relative wire positions, and configurable routine bodies. Identity computation must leave useful work to cache; recomputing expensive evaluations can defeat the optimization. Keep type-erasure and identity proposals provisional until their full obligations are justified.
Measure the optimization people will use
Check whether benchmarks include setup and allocation, measure their named operation, and reach the claimed hot path. Use instruction counts and whole workload timing for their respective questions; microbenchmark wins or noise do not establish fuse throughput. Qualify historical figures by rank and configuration.
Inspect collections, iterator cloning, copy-on-write, repeated evaluations, and hand unrolling. Identify where deferred tasks actually run concurrently and join. Check relevant features and worker counts. Preserve simple serial paths when parallel abstractions add allocation. Deterministic benchmark inputs are useful when the measurement requires them, not a universal rule.
Prefer concrete, legible APIs
Read signatures as a caller: prefer clear names, flatter arguments, and existing implementations. A wireless emulator may already provide native behavior for a driver-generic circuit. Require consumers for new traits, aliases, wrappers, and getters; semantic types help only when they clarify a role. Consider derive improvements before manual or unsafe boilerplate. Check proposals against actual driver bounds, gadget types, and returns.
Explain the reasoning the implementation cannot show
Make algebra, assumptions, and rationale auditable. Start with motivation, define conventions, and use a worked example when needed. Link to rustdoc; let guides add usage examples. Distinguish final stage/whole circuit, local/accumulated counts, child/parent, and application/shared stages. Recursion depth and security-model claims need arguments beyond implementation support.
Give precise, revisable feedback
When uncertain, briefly restate your reading and ask the specific question that would confirm or refute it. State demonstrated defects directly; reserve questions for unresolved uncertainty, and still check the relevant constraint. Give an equation, counterexample, signature, rewrite, or measurement that resolves each concern. Separate defects, open proof obligations, performance questions, and nits; a convention mismatch is not automatically a security bug. Report current locations and validation limits, revise mistaken arguments, and avoid finding quotas.
Consult references/evidence.md when provenance is requested or an ambiguity about applying a priority could change a finding. Historical examples are not current security verdicts and do not override current code or repository policies.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

