Let the AI interview you,
then make a plan.

Most advice about getting more from AI is advice about writing a better prompt. I want to make the opposite case. On real projects, the biggest gains I see don't come from a cleverer prompt. They come from two moves that happen before the real work starts. First, you let the AI interview you, so the goal is clear to both sides. Second, you make it write a plan and show you that plan before it runs. In late 2025 both moves got a name and a shared format. They're called skills. Here's how to use them.

The prompt was never the bottleneck

Watch how most people use AI at work. They type a request, read what comes back, sigh, and type a longer request. The mental model is that the prompt is the skill. Write the magic words and the machine behaves.

I've spent years designing AI systems for large companies, and that model has never matched what I see. The prompt is rarely where things break. Things break earlier, in the gap between what a person meant and what they actually said. Things break later, when the AI does a large amount of confident work in the wrong direction. The prompt is just the thin slice in the middle we happen to stare at. I've argued before that the collaboration, not the model, is usually what fails.

So the useful question isn't how to write a better prompt. It's how to close the gap before the work, and how to catch a wrong direction before it becomes a finished thing. Two habits do exactly that. The new part is that you no longer have to remember to do them by hand. You can package them.

What a skill actually is, in plain words

Start with the word, because it's doing real work now. In October 2025, Anthropic introduced skills as a way to extend what Claude can do, with the feature available across the Claude apps, Claude Code, and the API.^[1] A skill is a small folder with a file called SKILL.md inside it. That file holds plain instructions, and the AI loads them only when they are relevant.^[3] Anthropic's engineering team calls the idea progressive disclosure: the AI sees a short description all the time, and reads the full instructions only when it decides the skill applies.^[2]

Here's the plain version. A skill is a way to save a good way of working and hand it back to the AI whenever it fits. You write down the steps once. After that, you can type a slash command to run it, or the AI notices the moment and runs it for you.^[3]

Two things make this more than a convenience. First, the instructions cost almost nothing until they are used, so you can keep a deep library without slowing anything down.^[3] Second, the format is now an open standard. In December 2025, Anthropic published the Agent Skills specification at agentskills.io, and other tools adopted it quickly, from Microsoft's VS Code to OpenAI's Codex to Google's Gemini command line.^[4][12] A skill you write once isn't locked to one vendor.^[13] For a leader, that last point matters more than it looks. It means the way your best people work can become a shared asset instead of a habit that lives in one person's head.

Skill one: let the AI interview you

Now the first habit. Instead of you writing the perfect brief, you let the AI ask you questions first.

This sounds backwards until you see why it works. When a request is vague, the model doesn't stop. It fills the gaps with assumptions, and sometimes it invents details to make the answer look complete. Anthropic's own prompting guidance is blunt about this: under-specified prompts lead the model to guess, and a guess can turn into a confident wrong answer or a fabrication.^[5] The same guidance recommends the fix. When something is unclear, the model should say so and ask a clarifying question, and a good instruction reads like a short contract.^[5]Getting that right before any output is the heart of what Anthropic now calls context engineering, designing what surrounds the prompt rather than the prompt alone.^[6]

An interview skill turns that advice into a habit you don't have to remember. The SKILL.md says something like this: before you do this task, ask me up to five questions about the goal, the audience, the constraints, and what "done" looks like, then wait for my answers. That's the whole trick. You're forcing discovery to happen before output.

A concrete case. Ask for a customer feedback dashboard and a model will happily build one. But which feedback, and from where? What decision is the dashboard supposed to inform? Who opens it on a Monday morning, and what do they need to see in ten seconds? What counts as finished? An interview surfaces all of that before a single chart exists. If you've ever sat in a kickoff where the real requirements showed up three weeks late, you already know what this is worth. It's requirements gathering, the unglamorous skill good conversation designers and product people have always relied on. The interview makes the AI do that work instead of skipping it.

There's a quieter benefit too. The questions teach you. Half the time the AI asks something you hadn't settled in your own mind. You came in thinking the goal was clear. The interview shows you it wasn't. That's not the AI being slow. That's the cheapest moment you'll ever have to change your mind.

Skill two: make it plan before it acts

The second habit catches the other failure, the one that happens after the prompt. You make the AI write a plan and show it to you before it does anything.

Claude Code includes a built-in version of this called Plan Mode. In Plan Mode the AI is read-only. It reads your files, maps how they fit together, and writes out what it intends to do, but it changes nothing until you approve.^[7][8] You read the plan in plain language, edit it, push back, and only then say go.

Why does a checkpoint help so much? Because reviewing a plan is far cheaper than reviewing a finished thing. A plan is a paragraph. A finished thing is an hour of work pointed in the wrong direction. Anthropic's own guidance for Claude Code is direct about it: the explore and plan phases are the cheapest in effort and the most valuable in outcome, and skipping them is exactly when the AI builds the wrong thing, confidently.^[7]

This isn't only a coding trick, and it isn't a matter of taste. The research under it is some of the most cited work in the field. In 2022, Jason Wei and colleagues showed that prompting a model to lay out its reasoning step by step, a method called chain of thought, sharply improved results on hard problems.^[9] The whole industry took the lesson. Today's reasoning models from OpenAI spend internal tokens to think and plan before they answer.^[10] The model is doing privately, at machine speed, the same move you're doing out loud with a plan you can read. The difference is that your version keeps a human in the loop at the one moment when steering is cheap.

Why the two work better together

Run them in order and you get two gates before any work that is hard to undo. The interview sets the target. The plan sets the path. Nothing expensive happens until both are clear.

That sequence isn't something I made up. It mirrors the workflow Anthropic recommends for its own coding agent: explore, then plan, then act.^[7] It also mirrors how Anthropic describes building reliable agents in general, as systems that break a goal into clear, checkable steps rather than one large leap.^[11] When Anthropic's own teams describe how they work with these tools, the pattern that keeps showing up is structure first, output second.^[14]

For anyone responsible for risk, the two gates do something useful on their own. They produce artifacts. The answers from the interview and the approved plan are both things you can read, save, and point to later. When something goes wrong, you can see whether the goal was wrong or the path was wrong. Failure becomes locatable instead of mysterious. That's the same instinct behind treating evaluation as the product, not an afterthought.

What this means for leaders

If you lead a team putting AI into real work, here's how I'd translate all of this.

For customer experience, notice that you already want this discipline in your customer-facing AI. You want a support assistant that confirms what the customer means before it acts, and that doesn't invent a refund policy to fill a gap. The interview and the plan are the same discipline, turned inward and applied to how your own people work with AI. A team that clarifies and plans on its internal work builds better instincts for the customer-facing kind, the kind I write about in scaling LLM support to millions of users.

For governance and oversight, the approved plan is a control point you didn't have before. "Show me the plan" is a simple rule that a non-technical leader can require and an auditor can understand. It puts a human decision between the AI and anything that touches money, customers, or production.

For knowledge management, skills are the most practical answer I've seen to a hard problem: how do you keep the way your best people work after they change teams or leave? You write it down as a skill. Because the format is an open standard, that knowledge stays portable across the tools your org already uses.^[4] Industry watchers have framed the move as Anthropic trying to set common infrastructure the whole field can build on, which is part of what makes it safe to standardize on.^[15] This is content engineering by another name, the same structured writing I keep arguing is quietly becoming a required skill.

For organizational readiness, expect a new kind of role. Someone has to write the interview skill, decide which questions matter, and keep the plan templates honest. Call it a skill author or an AI workflow designer. It rewards people who are good at process and plain language, not only people who can code. If you're trying to move non-technical staff into AI work, this is a natural door, and I've watched that exact transition work when the operating rituals were designed with care.

What to do now, and what to stop

Concrete next steps, in the order I'd take them.

Start small this week. Write two skills. One interview skill that asks clarifying questions before a common task. One plan skill that requires a reviewed plan before any change to a real system. You don't need an engineer. A SKILL.md is a short text file.

Stop rewarding speed to first output. The fastest path to a wrong answer is to skip the questions and skip the plan. Praise the person who came back with five good questions, not the one who pasted a finished thing nobody asked for.

Make "show me the plan" the default for anything that touches a customer, a record, or production. Keep the casual, no-plan flow for throwaway drafts where being wrong costs nothing. Match the gate to the stakes.

None of this is exotic. It's the oldest advice in good work, dressed in new clothes. Get clear on the goal before you start. Look before you leap. The only thing that changed is that the AI can now hold those habits for you, ask the questions on cue, and wait for your nod before it moves. Use that.

Human review and AI assistance. This article was developed using AI-assisted research, analysis, and drafting workflows. A human reviewer evaluated the content before publication. Sources were reviewed for accuracy at the time of publication. While every effort has been made to ensure accuracy, readers should independently verify information before making business, legal, financial, regulatory, or technical decisions.