Iterate with Claude: how to give him feedback that really improves results

A digital marketing agency, eight people, demanding clients. They were working on the slogan for a new coffee brand. For forty minutes, someone was writing the same thing in Claude with slight variations: “Make it more creative.” «Another, but more emotional.» «A different one.»

Claude responded quickly. Too much. Thirty proposals later, no one was satisfied. Not because the sentences were bad, but because none of them fit withwhat reallythey were searching. The creative director stopped the process. He didn’t ask for another slogan. He explainedwhythe previous ones didn’t work. And at that moment, Claude stopped making noise and started tuning.

That’s iterating with Claude. It’s not asking for “another version” until something comes out that you like. It’s giving you the information you need to improve.

The first answer is not the problem

Many users judge an AI by its first response. It’s a mistake. In real work, no one expects a collaborator to get it right the first time when the assignment is complex or ambiguous. Claude works the same. The first answer is forsee how you understood the order. Not to close the job.

If something goes wrong there, it is not a sign of incapacity. It is information. Iterating with Claude is not admitting that the model is mediocre. It is used as any competent professional is used: with judgment,feedbackand direction.

Treating Claude like a brilliant junior

A brilliant junior has two characteristics: he understands quickly and needs judgment. If you just tell him “this is wrong,” he doesn’t learn anything. If you explainwhatis wrong andwhy, improves at high speed. Exactly the same thing happens with Claude. The quality of the feedback determines the quality of the iteration.

This is not anthropomorphism. It’s pragmatism. Claude processes concrete instructions much better than vague assessments. “Be more creative” does not mean anything operational. “The second paragraph sounds generic because it doesn’t connect to the client’s real problem” does give you something to work with.

Three ways to iterate (and they are not interchangeable)

Not all feedback does the same. There are three clear moves, and using the wrong one wastes time.

1. Regenerate: the simplest

“Try again.” It is useful when you are not sure what is wrong or when you are looking for quick variations. It is useful for exploring. Not to tune. Regenerating is rolling the dice: sometimes something better comes out, sometimes it doesn’t, but you’re not directing the process.

2. Nuance: the most profitable

“Keep this, but change that.” It is the most profitable feedback for production. You don’t invalidate all the work. You adjust a specific piece: the tone, the start, the focus. Claude responds very well when he knows what to preserve. To qualify is to say “this first paragraph works, the second loses focus, rewrite only the second attacking this specific point.”

3. Resume: the most powerful

You treat Claude like a person: you explain to him what problems you see, why they are problems, and what direction the solution should take. You can even ask him for the plan first before rewriting. That’s where the quality jump usually appears.

«Your answer has three problems: A, B, C. Before rewriting it, tell me in your words how you are going to solve each one. If the plan convinces me, execute it.

Reading that plan is where the shared criteria appears. You are not correcting blindly. You are lining up direction before executing.

Criticize the answer, not the model

There is a subtle but key difference. This doesn’t work:«This is wrong. Do it better.”This yes:«The second paragraph sounds generic because it doesn’t connect to the client’s real problem. Rewrite only that paragraph attacking this specific point.

Claude has no ego, but he has structure. The more specific the criticism, the better the next version fits. Vague feedback produces a vague response. Accurate feedback produces an accurate response. It’s not magic. It’s mechanical.

When to insist and when to start from scratch

There is a moment when iterating stops contributing. If you have given good feedback and the answer still doesn’t fit, the problem usually lies before: the initial prompt was confusing, the objective was not clear, or the assignment changed without saying it.

There is no need to insist there. It’s time to stop, redefine and start with a clean order. Insisting without criteria only tires the model and you. It’s like asking someone to fix a building with the wrong plans: no matter how much you adjust the windows, the walls are still incorrectly placed.

The clear signal: if after two iterations of concrete feedback the result does not improve, the problem is not in the answers. It’s at the starting point.

Iterating is not wasting time, it is refining the assignment

The agency in the example did not need 30 more slogans. I needed to express what they really wanted. When the director explained why the phrases didn’t work—”too generic for a cafe that positions itself as artisanal and local”—Claude’s next proposal was not another random variation. It was exactly what they were looking for.

Iterating with Claude is not a patch for poorly made prompts. It is the natural way of working when the assignment has nuances that are only revealed when seeing the first result. The trick is to iterate with direction, not blindly.

---