LevelUpWith

Explore / Delegating Real Work to Claude: Tasks, Files and Apps / Lesson 2

Lesson 2: The Plan-Act-Check Loop

Explains how Claude carries out a task by planning steps, acting with tools, checking results and repeating until done.

Free to read, no account needed. Part of Delegating Real Work to Claude: Tasks, Files and Apps.

By the end of this level you'll be able to

  • Describe the plan, act and check stages of a task in order
  • Read a task's progress view to see which step Claude is on and why
  • Recognise the points in the loop where Claude pauses to ask you a question

Why the loop matters

In Level 1 you saw that a chat produces an answer, while a task produces finished work. This level explains how Claude gets from your request to that finished work. It does not do the whole job in one pass. It repeats a short cycle of three stages: plan, act, check.

A comparison: a competent contractor redecorating a room first lists the jobs, then does one, then looks at the result before starting the next. If the first coat of paint is patchy, the list changes. Claude works the same way, and you can see it happening.

Once you know the loop you can:

  • Tell which stage a running task has reached.
  • Predict when Claude will stop and wait for you.
  • Step in at the right time instead of after the damage is done.

The three stages, in order

  1. Plan. Claude reads your request and works out what the job requires. It breaks a large job into smaller steps. For big jobs it can split the work into parts and run some of them side by side.
  2. Act. Claude carries out a step with a tool. A tool is anything that lets it do something rather than just write: reading a file, creating a document, running a calculation, searching, or using a connected app.
  3. Check. Claude looks at what the step produced. Did the file open? Do the totals add up? Does the result match what you asked for?

Then it loops:

  • If the check passes, Claude moves to the next step.
  • If the check fails, Claude revises the plan and tries another way.
  • When every step is complete, Claude delivers the outputs and reports what it did.

Example. You ask Claude to turn 40 expense receipts into one summary spreadsheet.

  • Plan: read every receipt, extract date, supplier and amount, build the sheet, total it.
  • Act: read the first batch of receipts.
  • Check: two receipts are unreadable scans. The plan changes: flag those two instead of guessing.
  • Act: build the sheet. Check: the total matches the sum of the rows.
  • Finish: deliver the spreadsheet with a note about the two flagged receipts.

Note the middle step. The check is where a good task differs from a careless one. Claude's own check does not replace yours, though. It can miss errors, so you still review the result. A later level covers that review.

Read the progress view

While a task runs, Claude shows progress indicators so you can follow along. The exact layout varies by device and is changing as Anthropic updates the app, so learn what to look for rather than where each item sits.

Do this each time you start a task:

  1. Read the opening plan. Check that the steps describe the job you meant. A wrong plan is cheapest to fix now.
  2. Find the current step. The progress view shows what Claude is doing at this moment, such as which file it is reading or what it is building.
  3. Read the reason. Claude usually explains why it is taking a step. If the reason surprises you, treat that as a warning.
  4. Watch for scope creep. Ask yourself two questions:
    • Is Claude opening files or websites I never mentioned?
    • Is the job growing beyond what I asked for?
  5. Act on what you see. You have three options:
    • Leave it alone if the steps look right.
    • Send a message mid-task to add information or redirect the work.
    • Stop the task immediately if something looks wrong.

You do not need to validate every individual action. Watch the pattern, not each keystroke.

Know where Claude pauses

A task is not fully hands-off. Expect Claude to stop and wait for you at these points:

  • Before or during planning, when the request is unclear. If something important is missing, Claude may ask a question, for example "Which quarter should the report cover?"
  • Before permanently deleting files. Claude needs your explicit permission. You see a permission prompt and must select Allow before the deletion happens.
  • Before other actions that need approval. How often this happens depends on your approval settings, which a later level covers.
  • When an action is blocked as unsafe. Claude looks for a safer route or asks you directly.

When Claude pauses, do this:

  1. Read the whole question or prompt. Do not approve by reflex.
  2. Check it against your original request. Did you ask for this?
  3. Answer specifically. "Use Q3 only" is better than "yes".
  4. If you are unsure, decline or ask Claude to explain first. Declining costs a minute; a wrong deletion may cost the file.

A paused task is waiting, not broken. It resumes when you answer.

Key takeaways

  • Claude completes a task by repeating three stages: plan, act, check.
  • A failed check sends Claude back to revise the plan; a passed check moves it to the next step.
  • The progress view tells you what Claude is doing now and why. Read the plan first, then watch for scope creep.
  • You can message Claude mid-task to redirect it, or stop the task if something looks wrong.
  • Claude pauses for unclear requests, for permanent file deletion and for other actions that need your approval.
  • Read every prompt before you select Allow.

A 6-question knowledge check follows this lesson in the app, free and without an account.

Keep going: levels 3 and 4 are free

Create a free account to open the next two levels, save your progress and pick up where you left off. No card needed.

Next: Level 3, Cowork Is Becoming Claude: Names and Plans.

Create a free account →See the whole course

More like this

  • Claude at Work: No-Code Skills for Knowledge Workers

    Learn to brief Claude clearly, check what it gives you, and set up Projects, Artifacts, Skills, memory, connectors and Research for real work in sales, marketi…

    Course · Claude · AI & Agents · Capable · 18 levels · 4 free · Professionally curated by levelupwith.com

  • Prompting Claude: Instructions, Context and Testing

    Write prompts that Claude follows reliably, manage what goes into its context window, and test and improve your prompts with evidence instead of guesswork. Eac…

    Course · Claude · AI & Agents · Capable · 20 levels · 4 free · Professionally curated by levelupwith.com

  • Finished Work with Claude Artifacts, No Code

    Brief, refine, brand, check, share and export finished designs, decks, documents, dashboards and motion pieces in Claude without writing code. Every level ends…

    Course · Claude · AI & Agents · Capable · 18 levels · 4 free · Professionally curated by levelupwith.com

  • Rolling Out Claude: An Admin's Plan from Pilot to Value

    Build a complete, defensible Claude rollout plan for your team or company, one decision at a time: structure and identity, access, governance, spend and visibi…

    Course · Claude · AI & Agents · Capable · 20 levels · 4 free · Professionally curated by levelupwith.com

  • Claude Code Day to Day: From First Prompt to Safe Habits

    Install Claude Code on your own machine and work through a small sample project with exact commands, expected results and an exercise in every lesson. You fini…

    Course · Claude · AI & Agents · Capable · 18 levels · 4 free · Professionally curated by levelupwith.com

  • Claude Code for Teams: Rules, Skills, Hooks, Plugins

    Build a complete, shareable Claude Code setup for your team, one working file per lesson: CLAUDE.md and path-scoped rules, skills, slash commands, subagents, h…

    Course · Claude · AI & Agents · Capable · 20 levels · 4 free · Professionally curated by levelupwith.com