Mastery Module 10 · Planning & Orchestration

Think Before You Prompt

Last reviewed

Beginner

What you'll learn

~15 min
  • Use a 60-second planning exercise before opening your CLI tool
  • Write a brief specific enough for the agent to plan from
  • Distinguish what to plan (the what/why) from what to delegate (the how)
  • Send a one-or-two sentence goal in plan mode and answer the agent's questions
  • Run the build loop: plan mode, the agent's questions, approve, build, check, and keep what worked in AGENTS.md
ℹBefore you begin

The planning half of this lesson needs nothing before it. The build loop in its second half builds on Module 9 (Scaffolding): you should be comfortable asking an AI tool to generate a project from a prompt before you reach it. If you skipped Module 9, read Let the AI Set Things Up first.

By the end of this lesson, you will be able to take any project idea — a data dashboard, a department website, an internal tool — and turn it into a specific prompt that can dramatically improve your first-draft quality. The secret is not better AI tools; it is better thinking before you type.

Mental model: the restaurant order

Think of prompting an AI tool the way you think about ordering at a restaurant. If you sit down and say “bring me food,” the kitchen has no idea what you want. You will get something, but probably not what you had in mind. However, if you say “I would like the grilled salmon, medium, with the roasted vegetables instead of fries, and the house vinaigrette on the side,” the kitchen knows exactly what to prepare.

AI tools follow your instructions as best they can, but they also infer missing details — so specificity reduces unwanted assumptions. Vague prompts produce vague results. Specific prompts produce specific results.

The most common mistake

Here is how most people use AI tools for the first time:

“Make me a website.”

The AI produces something. It is not what they wanted. They say “no, not like that” and get something else that is also wrong. After 20 minutes of back-and-forth, they are frustrated and the project is a mess.

The problem is not the AI. The problem is the prompt — and specifically, the thinking (or lack of it) that happened before the prompt.

The 60-second planning exercise

Before you open your CLI tool, spend 60 seconds answering these four questions:

1. What am I building?

Be specific. Not “a website” but “a portfolio site for a freelance photographer with a grid gallery, about page, and contact form.”

2. Who is it for?

This shapes everything. A tool for yourself can be rough. A tool for clients needs to be polished. A tool for beginners needs to be simple.

3. What does “done” look like?

Describe the end result. “When I open the page, I see a dark-themed grid of photos. Clicking a photo opens it full-screen. There is a navigation bar with three links. The contact form sends an email.”

4. What are the pieces?

Break it into parts. “I need: (1) a navigation bar, (2) a photo grid, (3) a lightbox for full-screen viewing, (4) an about page, (5) a contact form.”

💡Write it down

Jot your answers on paper if it helps you think. You do not need to turn them into a long prompt: a sentence or two is enough, and the agent will ask for the rest in plan mode (the build loop below).

Worked example: planning a real project

Let us walk through a complete example. Suppose you want to build an internal dashboard for your team that tracks project status.

Step 1 — What am I building? A project status dashboard that shows all active projects with their current phase, owner, and deadline.

Step 2 — Who is it for? My team of 8 people. They are not technical. It needs to be simple and visual.

Step 3 — What does “done” look like? When I open the page, I see a table of projects. Each row shows the project name, owner, status (a colored badge), and due date. There is a button to add new projects. Clicking a project opens its details.

Step 4 — What are the pieces?

  1. A header with the team name
  2. A table of projects
  3. Status badges (On Track, At Risk, Blocked, Complete)
  4. An “Add Project” form
  5. A detail view for each project
  6. Data saved to localStorage (the browser’s built-in storage, so entries survive a refresh)

Now squeeze that into a message of one or two sentences:

I want a simple dashboard for my team of eight to track project status. Each project shows an owner, a colored status, and a due date, and I can add new ones.

That took about a minute. In plan mode the agent will ask about the details, such as status names, colors, and where the data is kept, and you answer one question at a time.

🧬In Your Field: Biotechclick to expand

Lab inventory tracker. The same 60-second exercise works for research tools. Before prompting, answer: “I am building a chemical inventory page for our lab. It is for our 5 grad students who need to check stock before ordering. Done means a searchable table of chemicals with name, CAS number, location, quantity, and reorder status.”

🏛️In Your Field: Government Devclick to expand

Public records request tracker. Government teams can plan the same way: “I am building a FOIA request tracker. It is for our 3-person records team. Done means a dashboard showing open requests with requester name, date received, days remaining, and status.”

Why this works

You met vague-versus-specific first-hand back in Module 2. Planning is how you generate that specificity on purpose: the four questions force the detail into your prompt before you type it, so the first draft lands far closer to what you pictured. The anatomy of a strong instruction — context, examples, constraints — gets its own deep dive in Giving Good Instructions.

The 60-second mental check

Before you open your tool, answer three questions in your head. You do not need to write a document.

  1. What does it do? One sentence.
  2. Who uses it? You, your team, or the public.
  3. What does “working” look like? What you will see or click when it is right.

That is enough. The agent will ask about the rest, so you do not need to list pieces or choose tools. For example, a recipe app:

I want a recipe collection app for myself, where I can add recipes, search them, and tag them. When it works, I can find a recipe in a few seconds.

Send that in plan mode and answer what the agent asks. You are not skipping the planning. You are doing it as a conversation.

The build loop: plan, approve, build, check, keep

Your 60-second check is the brief, and this loop is how you hand it over for every build from here on. Switch the tool into plan mode first, on a turn of its own (a turn is one message you send):

/plan

Then send your one or two sentences on the next turn:

I want a recipe collection app for myself, where I can add recipes, search them, and tag them. Ask me what you need before you plan.
  1. Enter plan mode as its own step. Press Shift+Tab until the status bar shows plan, or send /plan (Module 6’s plan mode section): the tool reads and proposes before it builds; how firmly it waits differs by tool (the tool modes table).

  2. Let the agent interview you. It can read your folder, so its questions land on what your brief left open: what an empty recipe list shows, or whether tags are a fixed list or free text. Answer them. Questions here mean the loop is working; a long list of basic ones means the brief was thin.

  3. Approve the plan. Read it and ask for changes until it matches what you pictured in your 60-second check. Then accept it when the tool offers to proceed, or leave plan mode and tell it to build (the tool modes table says which your tool needs). Wait to approve until the plan matches what you want. Approving lets the agent change your files. Plan mode asks the agent to hold edits back, and how firmly it holds differs by tool; it is not a sandbox (a walled-off copy where changes cannot reach your project), so commit first, which saves a restore point you can return to (Git Workflow’s checkpoint: commit before the agent changes files).

  4. Build. The agent works through the plan you approved.

  5. Verify with the project’s checks. Start the app and click through what “working” looks like from your 60-second check. Then run the project’s checks, the commands that prove it still works, such as npm run build (and npm test once the project has tests), and read the diff (git diff, from Saving Your Work) to confirm it changed only what you asked. Ask the agent to run them and show you the result:

    Start the app in the background and give me the URL, then run the build and the tests if there are any. Show me the output of any command that failed.

    Prove It, later in this track, shows how to write those checks; until then the build command is enough.

  6. Keep what worked. When you corrected the agent, or found a command it should always run, ask for one line in the project memory file (Let the AI Set Things Up’s rule: commands and rules only):

    Add one line to AGENTS.md: show dates as YYYY-MM-DD. Then show me the file.

    (In Claude Code with a CLAUDE.md, name that file instead.) A whole prompt you find yourself sending again goes one step further and becomes a skill, a saved prompt file the agent runs by name (Giving Good Instructions’s rule: a reused prompt becomes a skill).

  7. Save the plan. Once the plan looks right, ask the agent to keep it in the project so you and the agent can both come back to it:

    Save the plan we agreed on to PLAN.md.

In browser chat, which has no plan mode, add “Propose a plan only. Do not write code yet. First ask me about anything the brief leaves open.” to your prompt.

See the Vibe Coding Reference for the full table of modes.

When planning is overkill

Not every task needs the full 60-second exercise. Simple, one-off changes like “rename the footer text” or “change the background color to #1a1a2e” can go straight to the AI without a plan. Use the planning exercise when you are:

  • Starting a new project
  • Adding a feature with multiple parts
  • Working on something you have not built before

If the task fits in a single sentence, just prompt it. Save your planning energy for the work that benefits from it.

The 80/20 of planning

You do not need to plan every detail. Let the AI Set Things Up’s rule holds here too: describe what you want and why, not how to build it.

Plan these (what/why):

  • What pages/sections does it have?
  • What can the user do?
  • What should it look like (general vibe)?
  • What constraints matter?

Don’t plan these (how — let the AI decide):

  • Which CSS properties to use
  • How to structure the JavaScript
  • What library handles the routing
  • How state management should work
📊In Your Field: MIS / Businessclick to expand

Business dashboard planning. When planning an analytics dashboard for your department, focus on what decisions the dashboard should support, not how the charts render. “Show monthly revenue trend, top 5 products by units sold, and customer acquisition cost” is a great plan. “Use D3.js with SVG elements and a responsive grid layout” is over-specifying the how.

🔧

When Things Go Wrong

Use the Symptom → Evidence → Request pattern: describe what you see, paste the error, then ask for a fix.

Symptom
The AI built something completely different from what I wanted
Evidence
My prompt was: 'Make me an app for tracking things'
What to ask the AI
"I want a habit tracker app. Each habit has a name and a daily checkbox grid showing the last 30 days. There's a button to add new habits. Completed days show green, missed days show gray. Save to localStorage."
Symptom
Plan mode answered with a long list of basic questions
Evidence
My brief was: 'Build a website', so every requirement was still open
What to ask the AI
"Build a personal portfolio website with: a hero section with my name and tagline, a projects grid showing 6 project cards with title/description/link, an about section with a bio paragraph, and a contact form. Dark theme, purple accents, modern and minimal. Before you plan, ask me only about what this leaves open."
Symptom
The result is too generic and looks like a template
Evidence
I asked for 'a business website' and got a generic Bootstrap page
What to ask the AI
"Create a website for a boutique coffee roastery called 'Morning Ritual.' Homepage has a full-width hero image area, three product cards (Single Origin, Blend, Seasonal), a story section about the roasting process, and a newsletter signup. Color scheme: dark brown (#2C1810), cream (#F5E6D3), accent gold (#D4A574). Typography: serif headings, clean sans-serif body text."

Practice: plan before you prompt

Run this on your machine

  1. Open your agent on a real project folder, such as your ai-class folder, then switch to plan mode inside it (Shift+Tab or /plan).
  2. Send a one or two sentence goal for something new you would like to add.
  3. Let the agent interview you and propose a plan.
  4. Pick one item in the plan and critique it. Is it too big, off target, or missing? Tell the agent what to change.
  5. Approve the plan only when each item is something you could check yourself.
KNOWLEDGE CHECK

What's the most effective way to start an AI-assisted project?

🔍Why specificity beats expertiseclick to expand

In practice, teams consistently find that the quality of the output depends more on the clarity of the input than on the user’s technical expertise. A biologist who writes “Build me a plate reader data visualizer with 96-well heatmap, dose-response curves, and EC50 calculation” will get a better result than a software engineer who writes “Make me a data analysis tool.” Domain knowledge — knowing what you need — is the real advantage. The AI supplies the technical how.

Key takeaways

  • 60 seconds of planning saves 20 minutes of frustration. Answer three questions in your head: What does it do? Who uses it? What does “working” look like?
  • Specific prompts produce specific results. Name features, describe layouts, specify colors and behaviors. The more detail you provide, the fewer assumptions the AI has to make.
  • Plan the what, not the how. Order the dish; let the kitchen choose the pans.
  • Keep the brief short. Send one or two plain sentences in plan mode, answer the agent’s questions, then ask it to save the agreed plan to PLAN.md.
  • Run the loop. Plan mode on its own turn, the agent’s questions, approve, build, run the project’s checks, then one line in AGENTS.md for what you learned.
  • Your domain knowledge is your superpower. You know what your field needs better than any engineer. That knowledge is what makes the prompt good.
Search lessons