Skip to content

How to ask your agent

Across this wiki, the setup step is usually "ask your agent." This page is the one-minute on-ramp for how to ask so it does the right thing the first time. The short version: say the outcome, name the thing, say where it lives — the way you'd brief a sharp colleague.

Time: 2 min to read; nothing to install. You'll need: Claude Code set up (~10 min once) — that's the agent these asks go to. Last verified: 2026-06-07

Before you begin

You don't have to learn commands. You describe what you want; the agent figures out the how and does it. The only skill is being specific — and that's mostly a habit, not a talent.

New to the whole idea? Start at Set up Claude Code, then come back. Everything below assumes you can type a sentence to a running agent.

The one sentence to remember

[Outcome you want]. [The specific thing — name the file/folder]. [Where it lives or what "done" looks like.]

That's it. A vague ask makes the agent guess; a specific one makes it act. Same idea Anthropic puts in its own docs: "Talk to Claude like you would a helpful colleague. Describe what you want to achieve." You wouldn't tell a colleague "make it better" and walk off — you'd say what and which one.

Vague vs. specific — the before/after

The fix is almost always the same three moves: name the thing, add the constraint, say what "done" looks like. Sharing-flavoured examples, the kind this wiki is about:

Vague — agent has to guess Specific — agent just does it
"share this" "put this folder in a private repo and invite sam as a collaborator"
"put it online" "deploy the site in this folder to Vercel and give me back the live link"
"send it to the team" "make a public repo from this folder so anyone with the link can read and copy it"
"make a link people can open" "publish the file report.html as a Claude Artifact, then check whether viewing it needs an account"
"let people edit it" "set up this budget tracker so my three teammates can each edit it, and tell me what each of them has to do to get in"

Notice the pattern: the outcome (share / publish / deploy), the specific thing (this folder, report.html), and a check or constraint ("private", "give me the link", "tell me what they have to do"). Steal that shape for any ask on this site.

Three habits that do the heavy lifting

  1. Point at the thing, don't describe it. "this folder", "the file donor-report.html", "the site I just built" — let the agent open it rather than retell what's in it. It reads better than your summary would.
  2. Ask it to hand back proof. End with "…and give me the link", "…and show me it worked", "…and tell me what the other person has to do." Now you can see the share actually happened, not just trust that it did.
  3. Course-correct the second it drifts. If it heads the wrong way, stop it and say so plainly — "no, keep it private" — rather than letting it finish. Anthropic's own advice: "Correct Claude as soon as you notice it going off track."

For anything bigger, split it into steps

If your ask has three parts, say three parts. The docs recommend this verbatim — "Break complex tasks into steps":

1. make a private repo from this folder
2. invite sam and priya as collaborators
3. give me the link to send them

The agent does them in order and you can see each one land. Much better than one tangled sentence it has to unpick.

If it doesn't work

  • It did the wrong thing entirely → your ask was probably ambiguous. Stop it (press Esc), and re-ask with the file named and the outcome spelled out. One specific sentence beats five corrections.
  • It asks you a clarifying question → good sign, not a failure — answer it. That's the agent being specific for you.
  • It guessed which file/folder → you didn't point at one. Tell it the exact name, or run the agent inside the right folder first (cd into it, then claude).
  • It keeps getting it wrong after 2–3 tries → the conversation is cluttered. Type /clear to start fresh, then give one clean, specific ask — Anthropic flags this as the single most common fix.
  • It won't do something risky-looking (deleting files, touching a password) → it's pausing for your OK on purpose. Read what it's about to do, then approve or redirect.

Watch / read

Best written intro: Anthropic's own Quickstart — Be specific with your requests is the shortest authoritative walkthrough; its "fix the bug" → "fix the login bug where users see a blank screen…" example is exactly this habit. Then Best practices — Provide specific context in your prompts is the fuller before/after table this page is modelled on. Both are primary docs.

Video: Your first Claude Code prompt — Claude (official channel), 2:27 — the right length for a first ask. Caveat: title and duration confirmed, but the transcript couldn't be pulled to verify the walkthrough on 2026-06-07; the two written docs above are the safer bet.

Sources