The latest articles on DEV Community by Gary Corriston (@gcorrist66). https://dev.to/gcorrist66 https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4034283%2F6cd8b5ac-3863-43af-aef1-fe0aaca2e7b6.jpg https://dev.to/gcorrist66 en Gary Corriston Fri, 17 Jul 2026 18:21:21 +0000 https://dev.to/gcorrist66/how-i-built-an-ai-review-skill-that-refuses-to-ship-bad-work-5ec4 https://dev.to/gcorrist66/how-i-built-an-ai-review-skill-that-refuses-to-ship-bad-work-5ec4 <h1> How I Built an AI Review Skill That Refuses to Ship Bad Work </h1> <p>The fastest way to create AI slop is not to use a bad model. It is to accept a polished first draft as proof that the work is finished.</p> <p>That failure shows up everywhere: a landing page uses an almost-correct brand color, a campaign invents a persuasive metric, or a Premium design is really a generic template with nicer spacing. An AI reviewer then compounds the problem with a confident “looks good.”</p> <p>I built <a href="https://github.com/gcorrist66/revue-proof-workflow-skill" rel="noopener noreferrer">revüe</a> to put a reproducible gate between the draft and the decision to ship.</p> <p>revüe is an MIT-licensed Agent Skill for Claude Code, Codex, and Cowork. It combines a review workflow with local validators, schemas, examples, and 108 automated eval cases. The central design choice is simple:</p> <blockquote> <p>Let the model handle judgment. Make deterministic rules executable.</p> </blockquote> <h2> The verdict is a contract </h2> <p>A revüe run ends with one of four verdicts: ship, ship with changes, caution, or block.</p> <p>In strict validation, ship is not accepted when required proof is missing or an audit fails. The model cannot simply write a reassuring paragraph and call the work ready. Its decision has to agree with the recorded evidence and audit state.</p> <p>Every non-ship verdict also needs an owner-tagged path to ship. This prevents review systems from producing an impressive list of concerns without closing the work.</p> <h2> Turn the brand system into a lock </h2> <p>For design and marketing production, revüe uses a design-system lock. It can define approved colors, font families, claims, banned patterns, and structural rules.</p> <p>The output validator inspects the actual HTML deliverable against that lock. Asking a model to remember the approved palette is useful, but it is not the same as checking the artifact it actually wrote.</p> <h2> Why the audit needed adversaries </h2> <p>The first version caught obvious violations. The red-team fixtures taught it to catch less obvious ones.</p> <p>The current suite includes near-miss colors expressed through HSL or modern RGB syntax, off-palette colors behind CSS variables, colors embedded in base64 SVGs, banned copy split with entities or zero-width characters, unapproved metrics, forged audit objects, hidden heroes, and placeholders hidden while guessed values are shown instead.</p> <p>Each discovered bypass becomes a permanent regression test. CI runs all 108 cases on every push and pull request.</p> <h2> Clean is not the same as Premium </h2> <p>One of the most useful failures was not a brand violation. It was a clean template declared as Premium.</p> <p>The page used approved colors and passed structural checks. It still lacked the craft expected of a flagship experience. That led to three explicit tiers:</p> <ul> <li>Standard: clean, on-brand, and lock-audited.</li> <li>Premium: everything in Standard plus an additional editorial craft profile.</li> <li>Custom: a human-directed bar with explicit signoff.</li> </ul> <p>Premium work is checked for machine-observable parts of that profile, including a full-bleed hero, display/body typography contrast, a repeating signature motif, and a persistent action treatment. These checks do not replace a designer. They prevent the system from equating compliance with craft.</p> <h2> Keep the trust boundary small </h2> <p>The validators use only Python’s standard library. They need no network connection, API key, or credential store. They read local artifacts and write local results.</p> <p>That was deliberate. A review gate is easier to trust when its dependencies and data boundary are small. It also makes the suite easy to run in CI or inside the self-contained installer.</p> <h2> Package one workflow across agent runtimes </h2> <p>The root SKILL.md contains the portable workflow. Supporting material lives in references, schemas and templates live in assets, and deterministic checks live in scripts.</p> <p>Claude Code users install the repository as a plugin marketplace. Codex users can install the skill manually or with the self-contained installer. The same instruction layer also works in Cowork.</p> <h2> What comes next </h2> <p>No validator catches every future evasion. The useful roadmap comes from concrete failures:</p> <ol> <li>Collect bypass reports as minimal fixtures.</li> <li>Add a failing eval before changing the validator.</li> <li>Make the smallest deterministic fix.</li> <li>Keep judgment-heavy checks visible as judgment, not fake certainty.</li> <li>Publish adoption numbers only when a public counter supports them.</li> </ol> <p>That last point matters because this project is also consulting proof for Corriston Consulting. The repository should demonstrate the operating principle we sell: AI speed is valuable only when the quality bar is explicit, testable, and honest about its limits.</p> <h2> Try it and try to break it </h2> <p>The code, examples, and eval suite are open under MIT:</p> <p><a href="https://github.com/gcorrist66/revue-proof-workflow-skill" rel="noopener noreferrer">https://github.com/gcorrist66/revue-proof-workflow-skill</a></p> <p>If revüe helps, star the repository. If it lets bad work pass, open an issue with the smallest artifact that reproduces the bypass.</p> <p>If your team needs a proof-first workflow around its brand system, approval rules, and client handoffs, that is the consulting work Corriston Consulting is built to do.</p> ai agentskills opensource python