Breaking a task down into three chained prompts

1Why three prompts beat a single one

Breaking down is the most profitable technique to learn after the four ingredients. Ask for three things in a single prompt: you will get one done well, one half done, and one ignored. Ask for the same three items in three successive messages, and all three come out usable.

Research measures this exact gain: splitting a problem into sub-questions handled in order takes a model from 16 percent to over 99 percent success on compositional tasks (Zhou et al., 2022).

The lesson's chain has three links. Prompt A produces a plan or a grid: the structure of a document, the criteria of a comparison, the columns of a table. You read it and correct it: this human validation is the move that changes everything.

Prompt B produces the content from the validated plan, with the source last and a rule for missing information. Finally, Prompt C reviews B's content against a short grid: does every statement have its page, its row, or its article? Do the bullets respect the length?

In Copilot Chat, the chain runs in a single conversation. Microsoft confirms that the conversation history remains accessible and that each turn keeps the context. Open one conversation per task, and paste the full prompt again when the exchange grows long rather than writing "like earlier".

For prompt A or C, enable "Think deeper" mode in the model selector when needed. This option hands the step to a deeper reasoning model, documented by Microsoft for analysis and verification.

One last setting, the one behind this lesson's booster: when you do not need to see the reasoning, ask Copilot to reason step by step internally and deliver only the result. When you need traceability, ask instead for the quotations before the analysis.

Tap to enlarge

2A forty-page report turned into eight slides

A project manager must draw eight slides from a forty-page report he attaches, for an executive committee.

Weak prompt:

Make me a presentation of this report.

Copilot proposes twelve slides with vague titles, figures without page numbers and a conclusion that is not the report's. Everything has to be redone, nothing can be checked quickly.

Strong prompt, in three stages:

Prompt A. From the attached report.pdf, propose a plan in eight slides for an executive committee that has fifteen minutes. For each slide: a title and the key message in one sentence. Wait for my validation before writing anything.

(validation: I merge slides 3 and 4, I replace slide 7 with "risks and expected decisions")

Prompt B. For each slide of the validated plan above, write three bullets of 12 words at most. Under each bullet, give the figure or the quotation from the report that supports it, with the page number. If no passage of the report supports a bullet, write "to be sourced" instead of filling in.

Prompt C. Review the eight slides. List every bullet without a page number and every bullet over 12 words. Rewrite nothing, only flag.

The plan arrives in eight lines and is corrected in two minutes. The bullets come out sourced page by page. The review flags two "to be sourced" bullets and one bullet that is too long, corrected by hand before formatting.

What changes: the plan is yours, not the model's, because you validated it before the writing phase. Every bullet points to a specific page, so checking takes five minutes instead of an hour.

The gap remains immediately visible ("to be sourced") instead of being filled by a plausible invention.

3Split a chase-up email into three prompts

Here is a fictional decision log from three milestone meetings of the "NovaTech SAS Portal" project, ready to be copied with no real data:

[Meeting 1 - 15 August 2026]
- Decision: Set up single sign-on (SSO) authentication. Owner: Alexandre. Due date: 30 August 2026. Status: Completed.
- Decision: Draft graphical design specifications. Owner: Beatrice. Due date: 25 August 2026. Status: In progress.

[Meeting 2 - 28 August 2026]
- Decision: Deliver approved Figma mockups. Owner: Beatrice. Due date: 5 September 2026. Status: Not started.
- Decision: Configure test database server. Owner: Alexandre. Due date: 2 September 2026. Status: Completed.

[Meeting 3 - 8 September 2026]
- Decision: User acceptance testing for order module. Owner: Camille. Due date: 7 September 2026. Status: Pending approval.
- Decision: Prepare internal training plan. Owner: Beatrice. Due date: 15 September 2026. Status: In progress.

The "Load exercise" button in the composer pre-fills the goal for Prompt A:

"Extract all actions into a table: action, owner, due date and status."

In a single Copilot Chat conversation, execute the chain in three steps:

• Prompt A (Extraction): Generate the complete table. Verify that each row matches the log above without omissions or hallucinations.
• Prompt B (Calculation as of 10 September 2026): Ask Copilot to calculate the late actions as of 10 September 2026 and the number of days late, strictly using the validated table from Prompt A.
• Prompt C (Targeted reminder): Request an 80-word reminder email per concerned owner, then check that every cited action exists in the table.

Self-assessment grid: (a) three prompts with a goal and a format each; (b) a written validation between A and B; (c) no email mentions an action absent from the table.

Open the prompt composer