I have lost more important exceptions to summaries than I care to count. A policy document would say a rule applied in most cases and then list three clear carve-outs. The summary would keep the rule and quietly drop the carve-outs. A research report would present a finding and then note the exact conditions under which it did not hold. The summary would report the finding as general. Each time the missing exception turned a usable note into a misleading one.
The problem is not that models are incapable of noticing exceptions. It is that most summary prompts never tell them the exceptions matter as much as the main claims. When the prompt only asks for a summary, the model reasonably optimizes for brevity and centrality. Exceptions lose.
The prompts below are built to reverse that priority. They treat the preservation of exceptions, limitations, and qualifying conditions as a first-order requirement rather than an optional nicety. They are short enough to test and specific enough to adapt.
The Core Problem These Prompts Address
Long documents often follow a predictable pattern: a general statement, followed by the conditions under which the statement does not apply, followed by the practical implications. Human readers know the middle part is load-bearing. Many summary prompts treat it as detail that can be compressed away.
A useful summary prompt for this kind of material must do three things:
Keep the main claims
Keep the exceptions and limitations attached to those claims
Make it obvious when an exception was present in the source but is missing from the output
The prompts that follow are variations on that requirement. I start with the smallest useful version and only add structure when the document type demands it.
Baseline Prompt (Start Here)
Summarize the following document.
Separate the output into two sections:
Main claims
Exceptions, limitations, and qualifying conditions
Keep every important exception with the claim it modifies.
If a claim has no stated exceptions, say so explicitly.
Do not drop limitations in order to make the summary shorter.
This is the version I use first on any long document that contains rules, findings, or recommendations. It is deliberately short. The explicit second section and the instruction to keep exceptions attached force the model to treat qualifying language as essential rather than optional.
Variation 1: Claim-Exception Pairs
Use when you want each claim and its exceptions kept together rather than in separate lists.
Summarize the document as a series of claim-exception pairs.
For each major claim, write:
Claim: [the main statement]
Exceptions or limits: [any conditions, carve-outs, or limitations that apply]
If no exceptions are stated for a claim, write “None stated.”
Do not invent exceptions or drop ones that appear in the source.
Variation 2: Priority on Exceptions
Use when previous summaries have repeatedly lost the qualifying language.
Summarize the following document with this priority order:
Preserve all stated exceptions, limitations, and conditions
Preserve the main claims those exceptions modify
Preserve only the additional context needed to make the claims and exceptions understandable
If you must cut material to keep the summary short, cut background and examples before you cut any exception.
Variation 3: Structured for Policy or Rules Documents
Use for terms of service, internal policies, regulatory summaries, or any document organized around rules and carve-outs.
Extract the key rules from this document.
For each rule provide:
The rule itself
Any stated exceptions or cases where the rule does not apply
Any stated conditions that must be met for the rule to apply
Present them in a simple list or table.
Do not generalize a rule beyond the scope the document gives it.
Variation 4: Research Findings With Scope
Use for papers, reports, or studies that present results with explicit scope limitations.
Summarize the main findings of this document.
For each finding include:
The finding as stated
The population, conditions, or scope to which it applies
Any limitations or contrary results the authors note
If the document does not state a limitation for a finding, write “No limitation stated in the source.”
Variation 5: Executive-Length Version With Safety Flag
Use when the summary must stay short and you still need protection against dropped exceptions.
Produce a short summary of this document (under 300 words).
Rules:
Every major claim must keep its stated exceptions or limitations
If an exception is cut for length, replace it with the note “[exception removed for length – see source]”
Do not silently drop qualifying language
This version makes the cost of compression visible. I use it when space is genuinely limited and I still need to know what was left out.
Variation 6: Side-by-Side Check Version
Use when you want a quick way to verify what the summary kept and what it lost.
First, list every explicit exception, limitation, or qualifying condition you find in the document.
Second, write a summary of the main claims that incorporates those exceptions.
If any exception from the first list does not appear in the summary, mark it as missing.
This is a diagnostic prompt as much as a summary prompt. It is especially useful when you are testing a new document type and want to see how easily exceptions disappear.
How I Choose and Adapt
I do not keep all six variations in active use. I keep the baseline and two others that match the documents I see most often. The rest stay in a reference note for the weeks when the usual versions are not enough.
When I adapt a prompt, I change only one element. If I need both a stricter length limit and a table format, I test them separately. Combining untested changes makes it impossible to know which part improved exception retention and which part introduced new compression problems.
A practical starting sequence:
Run the baseline on a document you already know contains important exceptions.
Check whether every exception you care about survived.
If some were lost, try Variation 2 (Priority on Exceptions) on the same document.
Keep the version that retains the exceptions with the least extra editing.
That single comparison usually tells you more than reading another list of summary templates.
What These Prompts Cannot Do
They cannot recover an exception that was never written down.
They cannot judge which exceptions are material to your specific decision.
They cannot replace reading the source when the stakes are high.
In every case the output is a draft summary. I read it against the original document before I rely on it. The prompts reduce the chance that exceptions vanish unnoticed; they do not eliminate the need for human review.
A Note on Length Versus Fidelity
Many summary prompts optimize for shortness first. These prompts optimize for fidelity to exceptions first and accept that the summary may be slightly longer as a result. In practice the length increase is usually modest, and the reduction in misleading omissions is worth the extra sentences.
If a hard length limit is non-negotiable, use Variation 5. It forces the model to leave a visible marker when an exception has been cut rather than letting the exception disappear without a trace.
Building Your Own Version
Copy the baseline and run it on two or three long documents that you know contain important qualifying language. After each run, note which exceptions survived and which ones disappeared. Those disappearances are candidates for a single new instruction. Add the instruction, test it against the baseline, and keep it only if exception retention improves without creating new problems.
The goal is not a perfect summary generator. The goal is a short, reliable method that treats exceptions as first-class content. Everything beyond that is optional.
Start with the smallest useful version—the baseline that forces a separate exceptions section and refuses to drop qualifying language for the sake of brevity. Test it on real documents. Add complexity only when the documents themselves prove you need it. The summaries stay longer by a little and safer by a lot. That trade-off has held up better than any shorter, more aggressive summary prompt I have tried.
No notes on this sheet yet.