Jason ColapietroFractional forward-deployed AI engineer. Founder of Suede AI, author, builder

Building products with AI coding agents

A useful AI coding workflow ends with a working change someone can check. These guides explain how to frame the work, keep parallel edits separate, and carry a change through review and deployment.

Start with an observable result

Describe the user action that should work when the task is finished. For example: a visitor opens a product page, understands the offer, follows its primary link, and reaches the right destination on a phone. Name the current failure and the expected result. Give the agent the repository and relevant page, then ask it to inspect the existing implementation. A brief such as ‘improve the site’ leaves the success condition open; a concrete journey gives both the agent and the reviewer something to test.

Establish which source is live

A folder name is not proof that it controls the public product. Match the repository to the hosting project, the production branch and the deployed commit. Read the local instructions and check for unfinished edits before changing files. When a task crosses several products, keep a small ledger with one row per site: source, affected page, intended change, verification and deployment state. This makes a missing surface visible before it becomes a forgotten one.

Keep the change small enough to explain

One task should have one coherent purpose. A content correction usually belongs in the page's source data or component; it should not quietly become a redesign or dependency upgrade. Put unrelated ideas in a separate task. For parallel work, use separate branches and working directories, and assign ownership of specific files. The goal is a diff where a reviewer can see the problem, the change and its consequences without reconstructing the entire project.

Check behavior at each boundary

A passing test answers only the question that test exercised. Check generated content after changing its source. Run the relevant repository checks. Open the page at desktop and phone widths, follow its links, and inspect empty and error states when they matter. After deployment, check the public address and the actual changed content. A successful build confirms that an artifact was produced; the live page confirms that visitors received the intended artifact.

Leave a useful handoff

Record what changed, why it changed, the checks that passed and any remaining uncertainty. Keep the exact commit and public URL alongside that note so a later session can distinguish source work from a shipped result. The stack guide explains the role of each tool; the operating rhythm turns the method into a daily review loop; the multi-agent guide shows when splitting a task helps. The contribution ledger linked from this site provides a separate, inspectable record of public engineering work.

What the method has shipped

The same way of working runs the whole Suede estate: more than 30 live sites indexed at hub.suedeai.ai, a family of iOS apps published under Jason Colapietro's Apple developer account, Suede AI Agent Studio and the IP Registry, and fixes merged into outside projects such as Jest, Backstage and the Polkadot SDK. This site is one of those surfaces.

Where the method is written down