Crypto Research Skill Creator

Crypto Research Skill Creator

Create research skills whose triggers are precise, conclusions are traceable to evidence, and cryptographic claims are not stronger than their proofs or sources.

Scope

Use this meta-skill to create, audit, or refactor a research skill. It defines the skill contract and quality gate; it does not perform the target research task unless the user separately asks for that task.

Keep the core portable across agent harnesses. Do not require a named tool, subagent model, hidden memory, proprietary metadata, or a particular directory outside the skill itself. A harness-specific adapter may be added only as an optional layer.

Required Inputs

Obtain or infer, without blocking on low-risk details:

• representative requests that should and should not trigger the skill;
• the research task, expected deliverables, and intended audience;
• required evidence sources and permitted verification capabilities;
• the relevant cryptographic domain and threat/security model;
• target harness constraints, if any.

If a missing choice would materially change correctness or scope, ask for it. Otherwise state the assumption and continue.

Success Criteria

A finished skill has:

• discriminating trigger metadata and explicit non-trigger boundaries;
• in-scope and out-of-scope work, required inputs, and observable outputs;
• success criteria, failure modes, stopping conditions, and uncertainty behavior;
• domain checks appropriate to its claims, parameters, and complexity measures;
• citation/version verification rules when external sources are used;
• valid relative links, no unfinished scaffold tokens, and no silent harness dependency;
• an audit result with blocking, major, and minor findings.

Failure Modes

Stop or downgrade the result when required evidence cannot be inspected, a citation cannot be verified, notation or parameters are under-specified, or the requested conclusion exceeds the available proof. Never fill gaps with plausible-looking references, theorem details, security reductions, parameter values, or novelty claims.

When a capability is unavailable, report exactly what was not verified and provide only the safe partial result. “Not verified” is a valid outcome; fabricated certainty is not.

Workflow

1. Select a mode: create, audit, or refactor.
2. Read design-principles.md to define the trigger, scope, contract, structure, and portability boundary.
3. Read crypto-research-conventions.md whenever the skill handles papers, proofs, reductions, security claims, ZKPs, lattices, parameters, complexity, citations, or novelty.
4. For creation, instantiate skill-template.md, remove irrelevant sections, and keep conditional detail in focused references.
5. For audit or refactor, preserve correct user choices and observable interfaces. Change only what a concrete finding justifies.
6. Apply validation-checklist.md. Resolve all blocking findings; disclose unresolved major findings.
7. Run python scripts/validate_skill.py PATH when Python is available. Treat it as a structural check, not a substitute for domain review.
8. Return the skill directory plus a concise audit report: verdict, evidence checked, changes made, unresolved risks, and unverified items.

Research-Integrity Invariants

• Label definitions, assumptions, conjectures, theorem statements, proof sketches, complete proofs, reductions, empirical observations, and engineering heuristics distinctly.
• A theorem's mathematical conclusion is not automatically an end-to-end protocol security claim.
• A proof sketch is not a checked proof; a machine-readable artifact is not machine-verified unless a checker was actually run successfully.
• State security notions, adversary powers, setup/model assumptions, parameter regime, and reduction loss before asserting security.
• Keep symbolic complexity, concrete cost, communication, memory, preprocessing, amortization, and failure probabilities separate.
• Verify bibliographic identity and claim support before citing a source. Never invent authors, titles, venues, identifiers, quotations, or theorem numbers.
• Treat novelty as a bounded search conclusion, not a fact. Use wording such as “no earlier result was found within the searched sources and dates,” together with search coverage and limitations.

Output Contract

For create, return a ready-to-copy skill folder and a validation report. For audit, return findings before suggested edits. For refactor, return the revised folder, a change summary, preserved behaviors, and remaining risks.

Use the examples only as structural guides:

• theorem-review
• literature-search