How to Use AI Coding Agents Effectively

Plan your sessions

• Break your task into a sequence of prompts, executed one by one
• Use plan mode: the agent reads, reasons, and proposes without writing code. Throwaway reads stay out of your real session.
• Track state in markdown files so the agent picks up where the last session left off

Context before code

• First: read the README, understand the test framework, reproduce the problem
• Lead with constraints and context, not implementation details
• All context up front is fine if it's reasoning about the plan, not executing

Parallelize your sessions

• Run each session in its own worktree or clone. Isolated working directories, isolated context windows.
• Each session has its own context window
• Three parallel sessions = triple your total context budget
• One concern per session

Smarter models are cheaper

A $0.50 prompt that gets it right is cheaper than five $0.05 prompts that each need fixing.

Don't optimize for token cost when the bottleneck is your time.

Manage your context window

Compaction loses dead ends, file reads, and failed attempts that still shape every response. Break problems down so you never need it.

Tell the agent how to verify its own work

• Example: after refactoring, write a script to check that only test cases calling suspending methods got runTest added and no annotations were removed
• The verification must be automated and deterministic
• Not "look at the code and tell me if it's right"

Invest in your local dev loop

• Your agent is only as good as your feedback loop
• Reproduce CI locally, run the service e2e locally
• Fast local feedback: lint, type check, unit tests, e2e, security scan
• If CI takes 20 minutes remotely, you're flying blind for 20 minutes

Don't rely on the model to remember your rules

• An instruction saying "never use rm -rf" can get ignored. An automated rule that blocks it cannot.
• Keep rule files small and scoped. A monolithic rules file fills context and later rules get ignored.
• Hooks, pre-commit checks, automated guardrails

Know when to bail

• Context corruption is irreversible. Failed attempts confuse every subsequent turn.
• Start a clean session. It takes 30 seconds.
• Bail signals: repeats itself, contradicts earlier decisions, suggests rejected fixes

Commit before everything

• Agents move fast and break things across many files at once
• Commit before every major agent operation
• git diff between commits is the fastest way to review
• Small commits let you revert a single bad step instead of an entire session

Work in agent-friendly languages

Know when NOT to use agents

• Pricing logic across products, admin permissions, infrastructure teardown. Anywhere subtle context spans many steps.
• High cost of failure + hard to verify = do it yourself
• Agents are confident and fast. If you can't verify quickly, that confidence works against you.