AI / Agentic
How I Use Claude, Codex, and Cursor Without Losing Control
By Vivek Jalondhara · 2 min read
Published Mar 2026 · Updated
Three tools, three jobs. The trick is knowing which one to reach for and where to draw the line.
I treat each tool as a specialist. Claude is my reasoning partner for design discussions, gnarly debugging, and writing precise specs. Codex and Cursor are for in-editor implementation where tight feedback loops matter most.
The guardrail is context discipline. I give the model exactly the files and constraints it needs — no more — and I keep changes small enough to review in one sitting. Big, vague asks produce big, vague messes.
Most importantly, tests are the contract. If the AI writes code, it also writes the test that proves it — and I verify that test can actually fail before I trust it.
Evaluate the output, not the tool name
The useful comparison is whether a tool can make a correct, reviewable change in the repository. A CRM permission change is a good example because the negative cases matter as much as the happy path.
Illustrative example · Review checklist · Simplified for explanation
Change: restrict editing of a protected CRM field.
Check: authorized users can still edit it.
Check: unauthorized API requests are rejected.
Check: hidden UI does not substitute for server checks.
Check: the diff does not expose customer data.- Same task
- Inspect proposed diff
- Exercise failure cases
- Accept or revise
Conceptual flow for the example above.
Tradeoff to keep in mind
Tool capabilities change. Keep evaluation criteria stable and choose based on the task and the review effort required.