TaskbyTask.Practical AI guides for work
ChatGPT, Claude, Copilot or Gemini · Practical walkthrough

Make an executive summary with an inspectable source trail

A summary should save reading time without spending the truth. "Trial complete in one office" has a way of becoming "live across the company" by the time it reaches the leadership team.

Task by Task6 min read + practice
A sculpted summary plaque sits above a transparent pedestal exposing two source blocks and one ghosted unresolved source.

What you'll make

a short executive summary where every claim carries a source reference, plus a claim map you can use to check the sentence that matters most.

What you'll need

your source notes with a short ID on each passage (S1, S2…), the reader and the decision they face, and a new chat in your work AI account (ChatGPT, Claude, Copilot or Gemini). About 15 minutes.

Example result.

The Bristol trial of the new travel booking service is complete: 35 bookings (S1).
Rollout to other offices hasn't been agreed (S1).
15 of 18 Bristol users rated it good or very good (S2).
Contract start is proposed for 1 February 2027, subject to legal review (S3).
Decision needed: whether to extend the trial to Leeds and London.

1. Ask for a short summary that keeps its conditions

A word limit pushes the AI to compress, and compression is where “proposed” and “one office” go missing. Ask for the conditions to survive, and for a source ID on every claim.

Write a short executive summary for [reader] about [decision] from these notes.
Start with the current state and the decision needed.
After each claim, give the source ID and quote the passage that supports it.
Keep proposed dates, conditions, sample sizes and anything still unknown.
Label any recommendation as a recommendation. If something can't fit accurately, cut a less important detail instead of its condition.

[reader] is who it’s for, such as “the operations director”; [decision] is what they need to decide.

Why this works: quoting the passage, not just citing the document, makes each claim checkable in seconds, and “cut a detail instead of its condition” tells the AI what to sacrifice.

What you should see: a summary of a few lines, each with a source ID, and a claim map quoting the passages.

2. Check the strongest sentence first

Find the claim that could change a decision, such as “live”, “all users”, “confirmed” or “agreed”, and open its passage. Compare three things: the status (done, proposed, pending), who or what it covers, and any condition. A real source link can still support a narrower claim than the summary makes.

Check the summary's strongest claims against their passages.
For each, compare the status, who or what it covers, and any conditions, in the source and in the summary.
Flag any claim that's broader than its passage, and give the shortest accurate wording.

Why this works: you’re checking a relationship between a sentence and a passage, which you can see for yourself. Asking whether the summary is “good” gets you reassurance instead.

What you should see: a short list of claims, each marked as matching its passage or broader, with corrected wording for the broader ones.

3. Read it as the person deciding

Read the corrected summary as the reader would, without the notes. Is the decision clear in the first lines? Does anything sound settled that isn’t? Keep the headline short: “Bristol trial complete; rollout not yet agreed” is accurate and still fits on one line.

Why this works: accuracy doesn’t need a headline full of caveats. One clear condition beats five hedges.

What you should see: a summary the reader could act on, with the conditions still in the sentences that need them.

Check the result

  • Every claim has a source ID, and the quoted passage says it.
  • Status words match the source: “proposed” is still proposed, “trial” is still a trial.
  • Sample sizes and scope are intact: one office is one office.
  • Recommendations are labelled as recommendations.

The usual failure: the summary widens a claim to make it sound finished. “Trial complete in Bristol” becomes “live across the company”, and “proposed, subject to legal review” becomes “starts on 1 February”. Spot it by checking every status word in the summary against its passage.

The summary says [overstated claim]. [source ID] says [what the source says].
Correct that claim and anything that depends on it, keep the summary short and keep the source reference.
Then check the other claims for the same kind of widening.

[overstated claim] is the sentence that goes too far; [source ID] and [what the source says] are its passage.

Worked example: a travel booking trial

Fictional example. An operations manager summarises a travel booking trial for the director.

S1 Trial completed in the Bristol office: 35 bookings in six weeks.
   Rollout to other offices not yet agreed.
S2 Feedback survey: 18 responses, all from Bristol. 15 rated it good or very good.
S3 Contract start proposed for 1 February 2027, subject to legal review.

First summary (wrong): “The new travel booking service is live across the company and staff rate it highly. The contract starts on 1 February.”

Repair: the repair prompt above, starting with “live across the company” and S1, then the check for the same widening, which caught the other two.

Repaired summary: the five lines in the example result at the top of this page.

  • “Live across the company” was the dangerous sentence. Read as written, it tells the director there’s nothing to decide.
  • “Staff rate it highly” turned 18 people in one office into “staff”. The repaired line gives the number and the office.
  • The contract date kept its condition. “Proposed, subject to legal review” is still short, and it’s true.

Try it

Open the passage behind the headline

Require the conditions behind each headline and repair unsupported compression.

1S1 Bristol trial only2S2 18 Bristol responses3S3 Start date proposed

A source link can support a narrower claim than the headline. Keep the status, the scope and the conditions when you shorten.

The example in text
  • S1 Bristol trial only
  • S2 18 Bristol responses
  • S3 Start date proposed

Task kit for this job: Download (ZIP) · Reusable instructions

Go deeper

Separate recommendations from findings

A recommendation can be useful, but it shouldn’t arrive looking like a fact from the source. Weak: “Extending to Leeds is the obvious next step.” Better: “Recommendation: extend the trial to Leeds for six weeks. Basis: S1 and S2.” The reader can see where the evidence stops and your judgement starts.

Turn the summary into a visual memo

When the relationships are easier to see than to read, the visual memo guide adds a source-linked diagram to the same claim map.

Keep a register for evidence you reuse

If the same sources feed several summaries, the evidence register guide shows how to store each passage once, and why finding a passage isn’t the same as showing it supports a claim.

Sources

Optional analytics

With your permission, Google Analytics uses cookies to measure visits and which guides people read. It stays off until you accept. Use Analytics choices in the footer to change your choice. Rejecting after accepting refreshes this page. Privacy details.