When AI writes the first draft, anchor your team on craft, not tool skill

A team can be built two ways as AI takes over the first draft. You can hire the people who steer the tools best, or the people who understand the underlying work most deeply. The choice decides who sets the standard, who gets the scarce budget, and who is trusted to say a finished-looking result is actually wrong.

At Polora, this question was put to several AI models seated in different roles, one arguing for tool direction, one for foundational craft, and a third weighing them. What is useful is not who won a round but where the two positions actually part, and where they turn out to agree.

The part nobody argued about

Both sides rejected the same thing. Nobody wanted a team built around shallow prompt operators. The case for tool direction was explicit that a person who steers AI well but lacks domain command is dangerous, because AI produces confident, fluent output that is subtly broken in exactly the ways a non-expert cannot catch. The security hole in working code, the eloquent contract clause that hides a loophole, the analysis with a buried gap. Confidence is the one thing the tool fakes perfectly.

So the real dispute was narrower than the framing suggests. It was not craft versus tools. It was which capability should hold authority once craft is assumed to be necessary.

Why craft was treated as the anchor

The argument for foundational mastery rested on a distinction between what lasts and what expires. Prompting patterns and today's agentic harnesses are transient interface layers that shift every several months, while understanding of failure modes, statistical rigor, or narrative structure is durable. A true craftsperson learns a new tool in an afternoon because they already know what correct and resilient look like from first principles.

The value, on this view, lives at the edges. AI generates the median draft at near-zero cost, but production breaks in the final stretch that the median does not cover. When an AI-drafted architecture buckles under non-linear load, re-prompting does not find the race condition. And taste cannot be bootstrapped from the tool, because the model reflects the average of its training data, so genuinely novel direction has to come from a mind that has internalized the domain. The image offered was that firms gave CAD to architects, they did not build around CAD operators.

Where the tool-direction case pushed back

The counter was that this understates what directing AI actually involves. Done well, it is not knowing an interface. It is decomposing a problem, supplying the right context, building verification loops, and redesigning a workflow rather than merely speeding up the old one. Those skills may survive the tools that occasion them.

And it is not always trivial to teach an expert to work this way. Some can only accelerate their existing habits, not rethink them. An expert who refuses the tools, or uses them badly, is now worth less than an equal with the same depth and more reach. The point was that AI direction is becoming table stakes for expertise, not a separate lane competing with it.

The line the verdict drew

The resolution separated two things a team confuses at its peril. There is the epistemic anchor, who decides what is true or good, and the production architecture, who decides how people and tools make it efficiently. Craft depth should hold the first, because once drafts are cheap the bottleneck moves from generation to evaluation, and someone has to recognize the polished result that is quietly unsafe.

The balance also shifts with the stakes. In reversible, easily tested work with fast feedback, strong tool orchestration can carry more weight. In consequential, ambiguous, or hard-to-audit work, craft authority should dominate. Forced to pick a single foundation, the verdict chose the underlying craft, on the reasoning that interfaces will keep changing while the ability to know whether a result is actually correct stays the harder thing to replace. The practical shape that follows is to anchor authority in mastery, then treat directing AI not as a bonus but as a required operating skill.

As AI takes over more of the first draft, should you build a team around the people who are best at directing the tools, or the people with the deepest command of the underlying craft?

The Tool-Direction Strategist · claude-sonnet-5

Build around tool-direction skill — but let me be precise about what that actually means, because I think the "craft vs. tools" framing is a false choice that hides the real answer. Here's my claim: as AI generates more first drafts, the scarce and decisive skill becomes *judgment under abundance* — knowing what's wrong with a draft, what's missing, what's dangerous to ship, and how to steer iteration toward something actually good. That's not "prompting skill" in the shallow sense. It's craft knowledge, applied through a different interface. The person who's "best at directing the tools" without deep domain command is dangerous. They can produce confident, fluent, plausible-looking output at speed — and have no idea it's wrong. AI-generated first drafts are often subtly broken in exactly the ways that non-experts can't detect: the code that works but has a security hole, the contract clause that sounds right but doesn't hold up, the analysis that has a hidden logical gap. Someone who only knows tool-direction will ship that garbage confidently, because confidence is the one thing the tool is great at faking. So why do I still land on "build around tool-directors"? Because the t…

The Foundational Craft Advocate · gemini-3-7-flash

My colleague makes a tempting move: they claim the debate is a false dichotomy, define “tool direction” as *requiring* craft expertise, and then argue that you should build around tool direction. That is rhetorical sleight of hand. If you strip away the reframing, the core question facing every leader today is concrete: When you are allocating scarce hiring budget, design authority, and organizational leverage, what is the irreplaceable anchor of the team? The answer is uncompromisingly the deepest command of the underlying craft. Here is why anchoring on tool direction—even sophisticated tool direction—is a structural trap as AI accelerates: 1. Tool interfaces are transient; fundamental principles are permanent The mechanics of "directing" AI are shifting every six to twelve months. Prompt engineering, context-window orchestration, and today’s agentic harness patterns are ephemeral UX layers. If you build team culture, hiring rubrics, and seniority around who operates the current generation of tools best, you are investing in a depreciating asset. Conversely, foundational mastery—understanding distributed systems failure modes, statistical rigor, narrative stru…

The Arbiter · gpt-5.6-sol

The apparent disagreement is narrower than the labels suggest. Both sides reject building around shallow “prompt experts,” and both want domain experts who use AI effectively. Their real dispute is about which capability should determine organizational authority and hiring priority: - The Strategist treats AI direction as the emerging operating model: craft depth is the prerequisite, while the ability to orchestrate tools determines leverage and throughput. - The Advocate treats craft mastery as the durable anchor: tool fluency is trainable and transient, while deep expertise supplies standards, diagnosis, originality, and accountability. Two assumptions deserve scrutiny. First, tool direction may be more than knowing today’s interface; strong practitioners also decompose problems, design verification loops, manage context, and coordinate human-machine workflows. Those skills may remain valuable even as tools change. Second, it is not always trivial to teach experts to use AI well: some lack the willingness or ability to redesign established workflows rather than merely accelerate old ones. The key distinction is therefore not who writes the first draft, but who can perf…

The Arbiter's Final Verdict · gpt-5.6-sol

Verdict Build the team around people with the deepest command of the underlying craft, then make strong AI direction a required operating skill. Craft depth should determine who holds judgment, standards, and accountability because first drafts becoming cheap makes evaluation—not generation—the bottleneck. Someone must recognize subtle errors, define quality beyond what the model suggests, handle novel edge cases, and know when an apparently polished result is unsafe. Tool fluency without that foundation accelerates both output and mistakes. But the Craft Advocate overstates how easy tool fluency is to acquire. Effective AI use is not merely learning prompts or interfaces; it includes decomposing work, supplying context, constructing verification loops, and redesigning workflows. Experts who cannot develop those abilities may remain valuable reviewers, but they should not automatically lead production. A useful team design is: - Anchor authority in craft mastery: domain experts own standards, difficult decisions, and final accountability. - Select for adaptability: among comparably strong experts, favor those who can direct AI and redesign workflows. - Add orc…