There is a gap between teams that use AI tools and teams that have genuinely reorganised around them. The first group integrates AI onto existing workflows; the second has changed how it thinks about delegation and what finished work means. A few patterns from Anthropic labs are worth examining because the distance between their practice and the industry average is large enough to be instructive.
From task delegation to goal-state prompting
The most important shift is also the hardest to realize: moving from step-by-step task delegation to expressing an end state, and letting the model identify trade-offs on its own.
Early AI-assisted workflows were naturally task-shaped — “write this function,” “fix this bug,” “summarise this document.” The model was an executor (a very fast one) of clearly scoped instructions. That still works, but it leaves most of the value on the table. The more useful pattern is to describe the outcome you want and the constraints that matter, then ask the model to propose an approach and flag the choices it had to make: the model becomes a collaborator.
This is not a prompt engineering trick. It requires actually knowing what you want at the level of goals rather than steps — a discipline in its own right.
Async and multiplayer, not individual sessions
Inside Anthropic, more than 60% of code is produced through an async delegation tool that lets teams assign Claude ongoing responsibilities via shared flows (es. Slack-like) — monitoring a feedback channel, picking up tasks as they arrive, handling routine API changes. The usage pattern is multiplayer and persistent, not a series of individual chat sessions that end when the window closes.
Seeing what your colleagues are delegating — and what the model produces — raises everyone’s sense of what is actually possible. A shared, visible delegation changes the culture of how the technology can be used, and not just the mechanics of it.
Most teams are nowhere near this. They are using AI tools in isolation, one person at a time, without shared visibility into what is being attempted or what is working. The bottleneck is not the model’s capability; it is the absence of a best practice.
The code review bottleneck is real
Even at a lab, the hardest constraint on speed is human review. This is not because reviewers are slow, but because large changes exceed what any human can fully hold in mind at once. The solution is structured artifacts that accompany a pull request: intent, trade-offs, and context, not just a diff. The reviewer gets a navigable explanation alongside the raw change.
A parallel practice: using the model to pre-review a pull request by asking the questions a careful human reviewer would ask, then making the final call as a human. So not automated review but using the model to identify the questions worth asking before a human spends time on them.
Ambition as a choice
There is a tendency to reach for AI tools for the tasks that feel safe: summarisation, first drafts, boilerplate. The more interesting question is what becomes possible once the constraints are removed.
As a concrete example: porting an entire Python codebase to TypeScript over a weekend — not as a thought experiment but as something that was done and worked — is a useful calibration. First-generation AI products were over-constrained in ways that made ambitious usage impossible. Those constraints are mostly gone. The limiting factor is whether a team is willing to try something that would have seemed unreasonable two years ago. The lesson? Be ambitious.
Simplification as a practice
One organisational detail worth borrowing: a Slack channel dedicated to actively removing features that have low usage, create cognitive overhead, or point in the wrong direction. The tendency in product development is to accumulate over time; this is a structural countermeasure to that tendency, provided there is an active review and the channel does not become a and a well-populated graveyard.
The underlying question applies at every scale: what exists that is not earning its complexity? Answering it honestly, and removing things, is harder than adding, which is why most teams never do it systematically.
The long game
The pace of change in AI is genuinely high — competitor moves that used to take a year now happen weekly. The teams that will still be effective in 3 years are the ones treating this as a 0n-going practice rather than a series of product launches to react to. That means actually resting, building the habit of naming what is hard before it becomes a crisis, and keeping perspective across the inevitable bad weeks.
These practices are not a formula but a set of practices. Which of them are already in place, and which represent the actual gap?
Text summarized and optimized using Anthropic’s models and reviewed by a human.