This lesson is being recorded

Every video of this course is being uploaded to my YouTube channel. Subscribe so you do not miss it.

Bruno AshfordBruno Ashford @brunoashfordSubscribe

Module 3 · Lesson 3.2 · 12 min

Explore, Plan, Execute, Review

Adopt the four-step loop Anthropic recommends for developers and use it for any business task.

Module 3
Your First Real Task

From testing the tool to delegating real work

Bruno Ashford
Bruno Ashford · Module 301 / 12
The shift

A chatbot waits. An agent goes and looks.

Claude Code works inside a folder on your computer.

  • Reads and searches your files
  • Plans before it acts
  • Asks before it changes anything
  • Checks its own work when asked
Bruno Ashford · Module 302 / 12
The loop

Explore, Plan, Execute, Review

  1. 1ExploreRead the data. No output yet.
  2. 2PlanApprove the approach first.
  3. 3ExecuteLet it work.
  4. 4ReviewIt checks, then you check.
Bruno Ashford · Module 303 / 12
A truth that hurts
Most bad AI output starts with asking for the answer before anyone looked at the problem.
Bruno Ashford
Bruno Ashford · Module 304 / 12
Safety first

Plan mode: think without touching

Press Shift+Tab until plan mode shows.

  • Claude reads and proposes
  • Nothing changes until you approve
  • Use it for anything touching real data
Bruno Ashford · Module 305 / 12
Messy spreadsheets

Five moves to clean data

Your original file is never touched.

  • Protect the original
  • Get a mess report
  • You set the business rules
  • Ask business questions
  • Reconcile the totals
Bruno Ashford · Module 306 / 12
Illustrative example, practice file

What a mess report finds

30rows in the sample file
4date formats in one column
3spellings of one customer
0changes to the original
Bruno Ashford · Module 307 / 12
Analysis vs. judgment

Ask for a decision, not a summary

Weak requestStrong request
What do you think?Choose between these 3 options
Summarise the dataCite numbers and source files
Give me insightsRisks and what proves us wrong
Write a reportOne page, recommendation first
Bruno Ashford · Module 308 / 12
The Delegation Brief

Brief it like a new hire

  1. 1GoalWhy the task exists
  2. 2ContextWhat a new hire needs
  3. 3InputsExact files to use
  4. 4OutputFormat, length, file name
  5. 5GuardrailsOff limits, and when to ask
Bruno Ashford · Module 309 / 12
Habits

Four habits of owners who delegate well

Small habits, big difference.

  • Ask it to look before acting
  • Read every permission prompt
  • Ask which files it used
  • Type /clear between tasks
Bruno Ashford · Module 310 / 12
Your project

A one-page diagnosis of your data

Real data, three recommendations, evidence for each.

  • Original file untouched
  • Totals reconciled
  • Every claim cites a number
  • An owner for every action
Bruno Ashford · Module 311 / 12
Recap
You do not need to be technical. You need to brief well and check the work.
Bruno Ashford
Bruno Ashford · Module 312 / 12
01 / 12

Use ← → to move. The marker highlights text and the pencil draws on the slide; navigation locks while you annotate. Undo removes the last mark, Esc exits.

1Explore: let it read without acting

Point Claude at the files and the question, and forbid changes. The goal is understanding, not output. This is where it catches the missing month or the duplicate customer.

Prompt for Claude
I want to understand why revenue dropped last quarter. First, read the files in the sales folder and tell me what data we have, what the columns mean and anything that looks wrong. Do not write any files yet.
2Plan: ask for the approach before the answer

Switch to plan mode with Shift+Tab, or simply ask for a plan. A good plan lists the steps, the files it will create and the checks it will run. Edit the plan like you would edit a junior's outline.

Prompt for Claude
Now propose a step-by-step plan to answer the question. Include what you will calculate, what files you will create and how you will check your numbers. Wait for my approval.
3Execute: approve and let it work

Once the plan is right, approve it. Claude will carry out the steps, asking for permission where it needs to write files. Let it run and resist the urge to interrupt every ten seconds.

Prompt for Claude
The plan is approved, with one change: split the results by product line as well as by month. Go ahead.
4Review: make it check its own work

Before you read the result, ask Claude to verify it. Totals should reconcile, every claim should point to a file, and anything uncertain should be labelled. Then you review with fresh eyes.

Prompt for Claude
Review your own work before I read it. Do the totals match the source files? Which numbers are you least sure about? List anything I should double-check by hand.
5Start fresh for the next task

When a task is done, type /clear before starting something unrelated. A clean conversation gives better answers than one full of old details, and it costs less.

Terminal
/clear

Here is a truth that hurts: most bad AI output is not the AI's fault. It is the result of asking for the final answer before anyone has looked at the problem. You would never let a new manager rewrite your pricing on day one without reading the numbers first.

Anthropic's own guidance for developers is to explore first, plan second, then build, then verify. It works just as well for a sales report, a price review or a client audit. Four steps, every time, and your hit rate goes up dramatically.

Key points

  • Explore first: no output until Claude understands the data.
  • Plan second: approve the approach, not just the answer.
  • Execute third: let it work without micromanaging.
  • Review last: make it check itself, then you check it.
  • Use /clear between unrelated tasks.

Avoid

  • Skipping straight to execute because the task feels simple.
  • Approving a plan you did not actually read.
  • Running five unrelated tasks in one long conversation.
  • Treating the review step as optional when numbers go to a client.
Real-world example: a roofing contractor

Dave's roofing company had a quiet spring and he assumed it was the weather. In the Explore step, Claude read two years of job records and noticed that quotes sent had held steady while quotes accepted fell by about a third. The Plan step focused the work on follow-up timing. The Review step caught that 11 jobs were logged twice. The real answer was slow follow-up after quotes, not rain, and it took one afternoon to find.

Bruno Ashford Subscribe