Build Cloudflare applications with AI assistance

Give AI current context, constraints, and an acceptance contract before generating, reviewing, and verifying a normal Web application.

AI ASSISTEDBEGINNER25 minutesOutcome: an executable AI development contract

AI-assisted development is a working method, not a product category. It can help you complete a SaaS, CMS, image utility, or PDF utility without any model call in the final application. What matters is that AI receives current Cloudflare documentation, project versions, runtime constraints, and acceptance criteria instead of guessing interfaces from memory.

Separate two ideas first

“Use AI to write the application” describes development. “Add AI capability to the application” describes a runtime product feature. This guide covers the first; the second stays optional.

The context packet

Environment facts

Node, Wrangler, Next/Fumadocs versions, package manager, and target runtime.

Resource contract

Existing D1, R2, KV, and Queue binding names, plus resources that do not exist yet.

Business boundary

Users, tenants, public routes, administration routes, input limits, and permissions.

Acceptance evidence

Required commands, HTTP states, failure branches, and rollback conditions.

Cloudflare Docs for Agents provides documentation entry points for coding assistants. Give AI the current primary pages together with the real project configuration. Do not provide only an old blog post, and do not let the assistant decide to create remote resources on its own.

Four-pass workflow

Ask AI to restate constraints

It should list confirmed facts, unknowns, and actions that mutate external state. If it misstates Pages, Workers, bindings, or a Wrangler field, repair the context before coding.

Define interfaces and failures first

Specify input, output, authentication, tenant boundary, maximum body, timeout, and downstream failure. Ask for a test matrix before implementation.

Generate in small units and verify locally

Complete one route or data-access layer at a time. Run lint, types, unit tests, and a workerd preview. Return exact errors to AI rather than saying that something “seems broken.”

Review external boundaries manually

Inspect secrets, token scope, remote migrations, DNS/WAF changes, cost, and deletion. AI may list commands, but it must not run remote writes without authorization.

AI workbenchHuman reviewed

Establish a development contract before generation

Use this with Cursor, ChatGPT, Codex, or another coding assistant.

Help me complete a Cloudflare Web application, but do not assume it is an agent or chat product. First inspect the package.json, Wrangler configuration, directory tree, and primary documentation links I provide. Then output only:
1. confirmed versions and runtime facts;
2. missing information;
3. business input, authentication, tenant, and data boundaries;
4. tests for success, 404, invalid input, unauthorized access, and downstream failure;
5. read-only local verification commands;
6. external state changes that need separate approval.
Do not generate code, create resources, or deploy until I confirm the contract.

Review focus

Surface	Frequent AI mistake	Acceptable evidence
Interface	Inventing an old configuration field	Current primary link and local type check
Security	Putting a secret in client environment	Server binding and clean secret scan
Data	Missing a tenant predicate or building SQL strings	Bound query and cross-tenant test
Runtime	Depending on local disk or a resident process	Real execution in workerd
Release	Treating upload, deploy, and migration as one action	Separate preview, version, and rollback steps

Completion checkpoint

The result should be independent of one AI vendor. If you change models, the same facts, boundaries, and tests still apply. If the assistant cannot identify unknowns or only covers the success path, development is not ready to start.

Next: build the complete application foundation.