The stack
Choose tools by the job they perform in the delivery process: understanding the code, making the change, checking it, and proving what reached the user.
The repository is the working record
Start with version control, the project's instructions and its existing scripts. Keep source, content, tests and deployment configuration together where the project already expects them. A generated website needs both the editable source and a reproducible build. A second copy of the same page in a scratch folder is useful for experiments, but it must not become a competing source of truth. Record which working directory and branch belong to the active task.
Coding agents need project context
Claude Code and OpenAI Codex can help inspect, edit and validate software. Their usefulness depends on the brief and the evidence available to them. Supply the failing behavior, relevant files, project commands and constraints. Ask for a proposed explanation when the failure is uncertain, then test that explanation. Save recurring project knowledge in the repository's instructions so the next session starts with the same operating rules instead of a vague recollection of an earlier conversation.
Skills package repeatable work
A skill is useful when the same job has a stable sequence, inputs and deliverable. A content review skill might specify where product facts live, how to verify links and what the final report must contain. It should also state when to stop, such as missing source material or an action requiring the owner's approval. Keep skills narrow enough to inspect. Browse the public skills catalog linked below to see the available workflows, then read the individual instructions before installing or running one.
Hosting is a separate verification step
The web estate uses Vercel, but a repository may contain several deployable products. Confirm the hosting project's root directory and branch before treating a merge as a release. Save the deployment URL and commit, then read the changed public page. For a static content change, compare the actual heading, body and canonical link. For an interactive change, complete the user action. Deployment logs help diagnose delivery failures; they do not replace checking the delivered experience.
Keep the toolset proportionate
A one-page copy fix does not need a new framework or an automation fleet. Use the repository's existing test runner and build pipeline, and add infrastructure only when a recurring failure justifies it. A practical starting set is version control, one coding assistant, the project's checks, a browser and a short delivery log. Agent Studio serves a different job: composing and reviewing agent workflows. Pick that workspace when the task itself is a repeatable workflow rather than a single code edit.
Google Cloud underneath
Suede has worked in partnership with Google Cloud since September 2025, with a dedicated account team and a migration of the stack to the Vertex AI and Gemini platform. The Vertex AI suite the stack moved onto includes Gemini, Vertex AI Agent Builder and Veo.
How this site's pages are checked
Changes to this site's generated pages merge behind one required check, ci-success, which passes only when the test suite and a production build pass. After each production deploy the live sitemaps go to IndexNow, and a daily canary checks every kept domain's home page, robots.txt and sitemap and confirms that each retired domain still answers with its permanent redirect.
The agent tools, in their own words
- Suede AI Agent Studio: AI Agent Builder & Workflow AutomationBuild AI agents visually for leads, invoices and reports. Test workflows, inspect every run, and publish paid x402 APIs when payments are enabled.
- Conduit: the execution layerThe proxy in front of every tool the agents can touch: who is calling, whether they are allowed, and whether it has been paid for.