AI Coding Workflow

Cursor, Claude Code, and Kiro aren't substitutes — they cover different task granularities. Sort out when to use which before mixing them, otherwise you'll just thrash between tools without shipping faster.

This page assumes you've installed at least one. If not, run through Install Cursor and the Claude Code Full Guide first, then come back for the team coordination side.

What each tool actually does best

In practice these don't conflict. Cursor wins when you're typing and need flow; Claude Code wins for "fix this bug across three services"; Kiro wins when you start a new feature ticket and want to nail the spec before any code is written.

Pick by task granularity

• One-line / single-function change → Cursor inline edit (Cmd+K). Faster than opening a chat.
• Logic inside one file → Cursor chat panel, reference the file with @.
• Cross 5+ files / needs test verification → Claude Code. Write a prompt that ends with "run bun test and verify."
• New feature from zero → Kiro writes the spec → Claude Code implements → Cursor polishes.
• Production incident debugging → Claude Code, because it can chain kubectl logs, psql, and grep through history in one session.

Stitching them into one day

# Morning: Kiro turns today's ticket into a spec
kiro spec create "add rate limit to /api/upload"
# → produces spec.md + tasks.md

# Late morning: Claude Code implements against the spec
claude
> implement tasks.md step by step, run tests after each step

# Afternoon: Cursor for UI polish
# Cmd+K "convert this toast to styled-components"

# End of day: Claude Code drafts the PR + changelog
> generate PR title and body from git diff main...HEAD

Rule of thumb: don't switch tools mid-task. Every switch means re-feeding context — pure token waste.

Measuring lift (not "AI feels fast")

Your tech lead doesn't care that AI feels productive. They want numbers. Track these four:

1. PR cycle time — hours from ticket open to merge, compared against the 30-day baseline before AI rollout
2. First-review pass rate — fraction of PRs that ship without revisions
3. Bug introduction rate — new bugs per week / PRs merged per week
4. Token cost — Cursor Pro $20/mo, Claude Code billed via API, Kiro tiered. Track per-engineer monthly spend.

Real numbers from JR Academy internal: PR cycle dropped from 2.3 days to 1.1 days after Claude Code rollout, but first-review pass rate fell in month one (engineers shipped more, reviewers fell behind). It recovered in month two. That lag is real — bake it into the rollout plan.

Common rollout mistakes

• No shared CLAUDE.md — each engineer's AI produces a different code style; review cost balloons. Commit a project-level CLAUDE.md into the repo.
• No one reads AI-written PRs — PRs balloon, reviewers rubber-stamp, three months later nobody understands the code. Cap PR size (e.g. +400/-200 lines).
• Token spend out of control — Claude Code defaults to Sonnet 4.5; a long session can burn $30+ a day. Set budget alerts; reserve Opus for hard problems.
• AI code without tests — if AI writes the implementation, AI must write the test. Otherwise the next AI edit will silently break behavior.
• Skills/commands stay personal — personal productivity ≠ team productivity. Promote high-value skills to repo-level .claude/skills/.

Next

• Vibe Coding in Practice — quality gates for natural-language-driven dev
• AI Rules Configuration — how .cursorrules / CLAUDE.md / Skills divide responsibility
• Claude Code Context Management — controlling token cost on long tasks