Summarise Without Losing the Point
Compress a long document to a tenth of its length while keeping what matters, then verify nothing important was dropped. Tests whether you can direct compression rather than just requesting it.
Why this one
"Summarise this" produces an average of the document — even coverage, no judgement. A useful summary is lopsided, because it knows what the reader needs, and that means telling the model who is reading and why.
Steps
Decide who the summary is for
A summary for someone deciding whether to fund something differs completely from one for someone implementing it. Write the reader and the decision down before you prompt.
Summarise against that purpose
Include the purpose in the prompt and require an explicit note of what was dropped — the omissions list is what makes the summary auditable.
PromptSummarise the document below for a reader who must decide [the specific decision]. Include only what bears on that decision. At the end, add "OMITTED:" and list the significant content you deliberately left out and why it does not affect the decision. Maximum [n] words. DOCUMENT: [paste]
Audit the omissions
Read the omissions list. If anything there would actually change the decision, the summary is wrong. This step is the whole reason for requiring the list.
Test it on the original
Take three questions the reader would ask and check whether the summary answers them correctly. Then check the same answers against the original document, because a fluent summary can be confidently wrong.
You should end up with
A purpose-built summary with an audited omissions list.
Done when
- The summary is explicitly built around one decision
- Nothing in the omissions list would change that decision
- Three reader questions are answered correctly against the original
If you want to go further
- Produce a second summary for a different reader and compare what each kept
Try next
Topics
Related across the site
Build a Personal Study Assistant
Turn a stack of lecture notes into a study partner that quizzes you, explains what you got wrong, and tracks which topics you keep failing. You will finish with a reusable system prompt you can point at any subject.
Context engineering
Context engineering is the practice of deciding what information enters a model's context window and in what order. As windows grew, choosing what to include became more consequential than wording the instruction.
Context window
The context window is the maximum number of tokens a model can consider at once — your prompt, any attached documents, the conversation history, and the response it generates. Exceed it and the earliest content gets dropped.
Fine-tuning
Fine-tuning continues training a pre-trained model on your own examples to specialise its behaviour. It is the right tool for teaching consistent style, format or classification behaviour — not for adding knowledge.
The 5-Part Prompt Framework — Claude
Every strong prompt combines five parts: Role (who the AI should act as), Context (background info it needs), Task (the specific ask), Format (how you want the output structured), and Constraints (length, tone, things to avoid).
10 Prompt Mistakes That Make AI Output Generic
The specific, fixable reasons AI output comes back bland — and what to write instead. Each mistake includes a before-and-after example you can apply immediately.