Talking to Claude

Once you’ve got the mechanics down, most of what makes Claude productive is how you talk to it: turn by turn, over a session, and over the life of a project.

A few techniques are worth naming explicitly. They’re not obvious, and they change how you work.

Rewinding

If Claude goes down a path you don’t like (wrong assumption, wrong direction, wrong tone), rewind to the point before the detour. Don’t argue with the response. Don’t try to coax it back. Just rewind.

In Claude Code, the command is /rewind. You pick which earlier turn to rewind to; the harness rolls the conversation back to that point and, if code changed in the interim, reverts the files from the checkpoint.

What’s actually happening: the harness holds the conversation as a transcript. When you rewind, the next turn is sent to the model with less context, specifically without the bad path. The model sees a cleaner history and responds to it fresh. (See the context budget for why that matters.)

This is also why rewinding covers both code and context:

• Code rewind: if Claude made changes you don’t want, files roll back too, via git or the harness’s checkpoint.
• Context rewind: the conversation state rolls back to the earlier point.

Reflex to build: when in doubt, rewind. It costs nothing, and the next turn gets the sharper response a cleaner transcript produces.

Branching

The same context can go down different paths. Once you’ve spent effort getting Claude into a particular state (a well-framed problem, a loaded plan, a clean spec), you can branch from there into multiple attempts without redoing the setup.

In practice:

• Reach a good “jumping off” point
• Capture it (a summary doc, a commit, or a note)
• Try one direction
• Rewind and try another

Especially useful for decisions that aren’t obvious: exploratory design, architecture choices, “X or Y” questions. One session, multiple branches, each informed by the same earlier context.

Synthesize

Freeform conversation is often the fastest way to ideate, but messy conversations don’t ship anything on their own. The move: talk freely for a while, then ask Claude to synthesize.

The pattern:

• Go back and forth on an idea without worrying about structure: what it is, what it’s not, what you want, what you’re unsure about.
• When you’ve said enough, ask: “Take everything we’ve been talking about and propose the shape of [the thing].”
• Read the synthesis. Rewind or correct where it’s off. Let it refine.

This is how a freeform dictation session becomes a spec, a PR description, an architecture sketch, or a first draft. You don’t have to organize your thoughts before you say them, as long as there’s enough for the synthesis to work from.

Failures as investment

When Claude gives you a response you don’t like, resist the reflex to lightly correct and move on. Instead: back up and invest.

Two levels:

1. Short-term: rewind, improve the prompt, try again. Get it right in this session.
2. Long-term: update AGENTS.md, a Skill, or another persistent doc so future sessions don’t have the problem. The correction becomes permanent.

Keep doing both until the response you want is the response you consistently get.

Every failure is data. Every correction is an investment in either this session’s prompt or the project’s long-term memory.

Ask why

Before you rewind and try again, try something counterintuitive: ask Claude why it did what it did.

The response is often genuinely useful. Claude can tell you which part of the prompt it was responding to, what assumption it made, or what trade-off it was trying to navigate. That explanation becomes a diagnostic: it tells you what to change before your next attempt.

The sequence:

• Claude gives you something you don’t like
• Instead of immediately rewinding, ask: “Why did you approach it that way?” or “What made you choose X over Y?”
• Read the explanation; it usually points directly at the prompt gap or the ambiguity that led to the bad output
• Then rewind, fix the prompt, and try again

This works because the model’s explanation is grounded in the actual context it was working from. That makes it a faster path to the right prompt change than trial and error.

The reflex

Over time, the default response to a bad output becomes:

Back up. What’s the smallest change (to this prompt, to AGENTS.md, to a skill) that would have made this go right? Make it. Try again.

Do it long enough and Claude stops making the same class of mistake. Not because the model improved, but because your context did. See managing context for the mechanics of the long-term side.

Resources

• Claude Code’s /rewind command and checkpoints (see the docs).
• Anthropic: Effective context engineering for AI agents