# Proof > Use for proof and tests, claims, invariants, behavior specs, edge cases, and evidence. - **Type:** Skill - **Install:** `agentstack add skill-kreek-consult-proof` - **Verified:** Yes — security-reviewed for prompt injection and unsafe behavior - **Seller:** [kreek](https://agentstack.voostack.com/s/kreek) - **Installs:** 0 - **Category:** [Agent Skills](https://agentstack.voostack.com/c/agent-skills) - **Latest version:** 0.1.0 - **License:** MIT - **Upstream author:** [kreek](https://github.com/kreek) - **Source:** https://github.com/kreek/consult/tree/main/consult/skills/proof ## Install ```sh agentstack add skill-kreek-consult-proof ``` Requires the [AgentStack CLI](https://agentstack.voostack.com/docs/cli). Works with Claude Code, Cursor, and any MCP-compatible agent. ## About # Proof ## Iron Law `NO ENGINEERING CLAIM WITHOUT A NAMED PROOF.` Prove every non-trivial claim before you call it done. A proof is a named check that would fail if the claim were false — point to that check, or mark the claim unproven. "The tests pass" doesn't count until you can say which test would break if you were wrong. ## When to Use As a completion gate, before any reply that says or implies the work is done, fixed, ready to commit, ready for a PR, or passing: - Name the claim and the check that backs it. If there is no check, say what is unproven and which evidence is missing — don't just write `unproven` and move on. As the main skill, when the task is the proof itself: - Writing or reviewing behavior tests for a feature, bug fix, refactor, flaky test, or any untested behavior. - Deciding what needs coverage and which boundary the test should enter through. - Turning an agreed spec, domain model, contract, or root-cause finding into Proof Contracts and runnable checks. ## When NOT to Use - Formatting, typo fixes, or file moves that change no behavior, data, or contract. - Mechanical refactors with no behavior surface — renames, file moves, comment-only edits — that tooling or a direct look at the result already confirms. (If the refactor changes something a caller can observe, you still need `proof`.) - Chasing a bug whose cause you haven't pinned down. Use `debugging` first; come back once you have a claim to prove. - Judging design, complexity, naming, or structure. Use `code-review`. Reach for `proof` only when the open question is whether the evidence is enough. - Load testing, profiling, or benchmarks. Use `performance`. - Setting up the test runner, linter, or typecheck baseline. Use `scaffolding`. ## Where Proof Enters Test at boundaries. A boundary is any point where one part of the system hands data to another and the receiver expects a certain shape — between modules, layers, or processes. (This skill also calls these handoffs.) Put the proof where the data's shape, value, state, or error visibly changes. A good boundary test passes against any implementation that keeps the same contract, so it survives refactors. The outermost boundary is the one your caller sees: the HTTP endpoint, CLI, UI, or public API. It always counts, because its behavior is what the user actually depends on. Boundary tests usually exercise the helpers beneath them for free. Add a separate unit test only for non-trivial pure logic — not for every function a boundary test already drives. ## Core Ideas 1. Every claim owes a check. If you can't point to one, mark the claim `unproven`. Silence is not proof. 2. You don't have to write the test first, and you don't always have to write one. Behavior changes need a runnable check; mechanical edits, prose, and facts the tooling already guarantees do not. Never test static text that only changes when someone hand-edits the file. 3. Match the proof to the claim. A typo fix needs nothing. A new endpoint needs a contract test. A subtle bug fix needs a regression test that fails before the fix and passes after. 4. "Done" is a claim too. A green check counts only when it proves the latest request was met. 5. Write the proof so it teaches the behavior. The next developer should read it and learn how the system is meant to work, not just see a checkmark. 6. Different claims need different evidence: data claims need an invariant; behavior claims need a boundary check; bug fixes need root-cause evidence and a regression test; refactors need the same behavior before and after. 7. One behavior per test. Use the real collaborators on the inside; mock only true system edges such as the network, clock, or filesystem. Don't test the framework, the language, or static copy unless your code makes it a contract. 8. A flaky test is a bug — in the test, the code, or the environment. Fix it; don't bury it under sleeps or retries. ## Proof Contract For each non-trivial claim, write down these five things. Filling them in is what "naming a proof" means, and other skills (`domain-modeling`, `refactoring`) hand work back here expecting it. - **Claim** — the behavior, invariant, contract, or root cause you assert. - **Data invariant** — the shape, state rule, or type that makes bad states impossible, or at least visible. - **Boundary** — where the claim becomes observable, i.e. where the check enters. - **Check** — the runnable test or command that would fail if the claim were false. - **Evidence** — what you actually saw: the command and its output, the test name and pass/fail, the artifact you inspected, or a plain reason you couldn't run it. ## Workflow 1. List the claims this change makes or relies on. Keep the ones a caller or user can observe, external contracts, domain invariants, "still behaves the same" refactor claims, and real error cases. Drop imagined edge cases, framework guarantees, and language behavior. 2. Fill a Proof Contract for each remaining claim before you call the work done. 3. Tie every named requirement to the artifact that satisfies it and the check that proves it. A green command isn't proof if it never runs that artifact. 4. For data, config, wiring, generated output, or documents, prove them the way the system uses them — run, load, parse, render, or inspect — not by asserting their literal text. 5. For a removal or replacement, prove the behavior that remains, not the thing you deleted. Don't write tests for ghosts; confirm the old code is gone with a search instead. Exception: if removing something now returns an explicit rejection (404, 410, a deprecation error), test that rejection — it is new behavior. 6. Pull in only the one reference you need: - `references/data-shape-boundaries.md` — worked boundary examples: pipelines, parsers, validators, middleware, sans-IO protocols, and functional-core/imperative-shell splits. - `references/recipes.md` — when the proof shape is specific to a domain. - `references/removals.md` — for removals and replacements. - `references/test-theater.md` — when a test asserts how the code is built instead of what it does. 7. When a claim needs a test, name it in the caller's words and assert the result the caller sees. 8. Use the narrowest check that proves the claim. Start with a single test by name or line; fall back to one test file; run the whole package or suite only after the narrow check is green and you suspect wider breakage. ## Before Saying Done 1. Re-read the latest request and any corrections. State what "done" means in the caller's words. 2. Look at any file or artifact the request named, as it stands after your last edit, and confirm your proof actually reads or runs it. 3. Run or inspect the check you picked in the workflow — freshly, not from memory of an earlier run. 4. Say where things really stand: proven, partly proven, blocked, or unproven. ## Verification - [ ] Every non-trivial behavior, invariant, contract, root-cause, or refactor claim has a Proof Contract. - [ ] A check enters at each boundary where the data's shape, value, state, or error visibly changes, plus the outermost caller boundary — and each would fail if the claim were wrong. - [ ] Test names and assertions describe what the caller observes, not private methods, call order, framework behavior, or static text that never varies. - [ ] Mocks sit only at true system edges, or carry a written reason. - [ ] Tests pass in any order and don't depend on sleeps. - [ ] Every file, script, config, command, or document the request named is tied to an artifact and a check. - [ ] Smoke checks, helper-only checks, and proofs of nearby behavior are flagged as partial — they don't count as acceptance. ## Tripwires - A green check that never touches what you changed (or only hits a helper or nearby behavior) is partial, not proof. Run the check that would fail if this change were wrong. - About to run the whole suite for a one-line edit? Run the single relevant test, line, or file first. - If a broad suite fails on unrelated drift, switch to the targeted check and report the broad failure separately. - If unit tests would just restate helpers a boundary test already covers, skip them and test the contract instead, so refactors don't have to rewrite proof. - When static types seem to already prove it, name the invariant or boundary behavior the types don't cover. - When the practical check is manual, record the command, the output, and the claim it proves. Common for config, build wiring, and generated files. - Rewrite or delete tests that assert how the code is built instead of what it does. See `references/test-theater.md`. ## Handoffs - `specify` — turn an agreed ADR, RFC, spec, or note's proof obligations into Proof Contracts before you claim done. - `domain-modeling` — shape the invariants and make invalid states impossible to represent. - `debugging` — when the proof hinges on root-cause evidence. - `api` — when the claim is a public contract. - `refactoring` — for before/after behavior evidence, and when a test that's hard to place is telling you the code is tangled and should be simplified first. Add `architecture` when the tangle crosses module boundaries. - `error-handling` — when the error shape, message, or recovery the proof must assert isn't settled yet. That skill owns the contract; this one owns the proof. - `security` — when the proof needs abuse cases or trust-boundary checks. ## References - Worked boundary examples (pipelines, parsers, validators, middleware, sans-IO, functional core): `references/data-shape-boundaries.md`. - Proof recipes by claim type: `references/recipes.md`. - Removals and replacements: `references/removals.md`. - Test-theater traps: `references/test-theater.md`. - The `consult` Pi package ships a `/proof` runtime command that runs a red-green-refactor cycle when behavior tests are the right tool. Its output counts only toward the claims it actually covers. ## Source & license This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code. - **Author:** [kreek](https://github.com/kreek) - **Source:** [kreek/consult](https://github.com/kreek/consult) - **License:** MIT Install and usage instructions live in the source repository linked above. ## Pricing - **Free** — Free ## Security capabilities Automated source analysis of v0.1.0 — what this tool can access: - **Network access:** no - **Filesystem access:** no - **Shell / process execution:** no - **Environment & secrets:** no - **Dynamic code execution:** no *"Yes" means the capability is present in the source — more access means more to trust, not that it is unsafe.* ## Versions - **0.1.0** — security scan: passed — Imported from the upstream source. ## Links - Listing page: https://agentstack.voostack.com/l/skill-kreek-consult-proof - Seller: https://agentstack.voostack.com/s/kreek - Browse the marketplace: https://agentstack.voostack.com/browse --- Listed on AgentStack — the marketplace for AI agent skills and MCP servers. Every listing is security-reviewed. Creators keep 70%.