Iterating & refining

The first answer is a starting point, not the finish line. Skilled users treat a chat as a collaborative editing loop.

The refine loop

1. Get a draft. Don't over-engineer the first prompt; get something on the page.
2. React specifically. Say exactly what to change.
3. Repeat. Two or three rounds usually gets you to great.

Good start. Now:
- Cut the intro paragraph
- Make the tone more direct
- Add a concrete example in point 3

Give feedback like an editor

Vague feedback gets vague improvements. Be precise about what and why:

Weak feedback	Strong feedback
"Make it better"	"Tighten it; aim for half the length"
"I don't like it"	"Too formal; write like I'm talking to a teammate"
"Add more"	"Add a section on risks, with 3 bullets"

Steer, don't restart

If Claude drifts, you usually don't need a brand-new prompt; just course-correct:

You changed the structure I asked for. Go back to the 3-section
format from before, keeping the improved wording.

Branch instead of overwriting

When you have a version you like but want to explore, ask for alternatives so you don't lose the good one:

Keep this version. Now also give me a bolder, riskier variation
as "Option B" so I can compare.

In claude.ai you can also edit a previous message to fork the conversation and try a different direction.

Save what works

When a prompt produces consistently great results, save it. That's the seed of a reusable template (see Prompt templates or even a Skill.

When to start fresh

Iterate within a chat when you're refining one thing. Start a new chat when:

• You switch to an unrelated task.
• The conversation has gotten long and Claude seems to "lose the thread."
• Early mistakes are anchoring the responses.

Long threads cost more and can muddy focus; a clean start is often the fastest fix.

Try it

Next time an answer is "almost right," resist re-typing the whole prompt. Give three specific edits instead. Notice how much faster you converge.

Next: the common pitfalls that trip people up.