How to Actually Work with AI
Most teams use AI like a search engine. Here's how to use it like a senior engineer.
Most engineers use AI the same way they used Stack Overflow — paste a question, grab an answer, move on. That gets you 10% of the value. The teams pulling real leverage out of AI tools treat them more like onboarding a new engineer: give them context, set expectations, verify their work, and close the feedback loop.
Context is the job
AI models know nothing about your codebase until you tell them. The teams that get the most out of these tools invest in making that context cheap to provide: CLAUDE.md files that explain project conventions, annotated indexes that surface the right files, and a consistent structure so the model isn’t guessing.
Think of it as documentation that pays you back immediately instead of six months from now.
Encode your preferences once
You shouldn’t have to explain your coding standards on every session. Write them down once — preferred patterns, error-handling conventions, what “done” looks like — and scope them appropriately: global defaults, repo-level conventions, project-specific context. The model follows a behavioral contract; you write the contract.
Repetitive workflows deserve the same treatment. If you’re walking the model through the same process more than once a week, turn it into a reusable skill it can invoke on its own.
Verification unlocks autonomy
The more the model can check its own work, the more you can step back. Wire up linters, formatters, and type checkers so the model gets deterministic feedback without waiting on you. Give it access to run tests and inspect output. Cheap, fast feedback loops are what turn an AI assistant into an AI collaborator.
Without them, you’re babysitting. With them, you’re directing.
Delegate outcomes, not steps
There’s a big difference between “write this function” and “implement this feature — here’s what it needs to do and how we’ll know it’s working.” The second prompt treats the model as an agent, not a typist. That shift in framing changes the quality of what you get back.
For larger work, parallel sessions in separate worktrees let you run multiple tracks without stepping on yourself.
Close the loop
The payoff compounds if you treat each session as organizational knowledge, not a one-off. Share useful patterns with the team. Mine transcripts for missing config or broken verification steps. Prune rules that conflict or drift out of date.
What you’re doing, session by session, is training a collaborator. That’s worth being deliberate about.