How well does your team work with AI?

The AI sounds equally sure when it's right and when it's wrong. Treat confident answers as guesses until you've checked them.

Example: The AI confidently names a command-line flag that doesn't exist on your platform. The team builds tooling around it. Hours lost.

Spot the moment the AI is agreeing too eagerly, inventing an API, or quietly dropping a constraint — and redirect it instead of absorbing the bad output.

Example: The AI proposes an architecture that drops a constraint from earlier in the session. You name what it forgot and ask it to try again.

The AI doesn't know your codebase quirks, last week's decisions, or your customer's environment. Tell it, instead of waiting for it to ask.

Example: A test fails because of a project-specific fixture. Asking "fix the test" won't work. Attaching the fixture file will.

Know when to think out loud with the AI and when to authorise it to act. Both extremes fail — endless dialogue, or executing too early.

Example: On a refactor, talk impact first, then say "go." Don't open with "go."

Break a vague ask into a sequence of small, specific questions the AI can actually answer well.

Example: "Build a CSV importer" becomes "agree on the schema, then validation, then persistence, then errors" — four sharp asks instead of one fuzzy one.

Know when to refine the current attempt and when to scrap it and re-prompt from a different angle. Refining for too long is the most common failure.

Example: The AI's first regex is structurally wrong. Don't ask it to "fix" — restart with explicit specs for each capture group.

Pick the right model, the right context to attach, and the right rules for the task. Top-tier for everything wastes money. Mid-tier for everything wastes the leverage.

Example: A fifty-file refactor wants a top-tier model with full context. A variable rename wants a small model with one file. The same setup for both wastes one or the other.

Throw out two hundred lines the AI just wrote the moment you realise the premise was wrong. Don't defend the work because the AI already did it.

Example: Halfway through, you spot a wrong assumption. The right move is "stop, back up." The wrong move is "let's patch around it."

Treat every AI answer as a guess until something independent confirms it — a compile, a test, the actual docs, a second source.

Example: The AI says the SDK has a batchUpload method. You check the docs before you write the call, not after the test fails.

Stay steady when the AI mis-fires and stay critical when it nails one. The last interaction shouldn't decide the next.

Example: After three failed attempts on a parser bug, you re-frame the question instead of concluding the AI "can't do" parsers.