You pasted three paragraphs of context — company background, the team setup, the current project status, why you are even doing this thing — and ended with “so, can you draft the launch email?” The model responded with a 400-word summary of your company. Not an email. Not a draft. A summary. The task was technically in there, but it was the 47th sentence of a 50-sentence prompt and the model lost track. When background dominates the prompt, the model interprets the request as “talk about this background” because that is what would happen if a person sent the same wall of text to another person.
This page walks through why long context buries the task and how to structure prompts so the task is unmistakable even when the context is necessary.
Common causes
1. Background paragraph 1 reads like an essay opening
If your first 200 words are “Our company was founded in 2019 with a mission to…”, the model anchors on “this is a profile of company X” and produces something matching that anchor. The actual task — buried later — does not change the anchor.
How to spot it: first 3 sentences of your prompt are descriptive prose, not imperative.
2. Task stated once, in prose, late
A single sentence — “can you draft the launch email?” — at sentence 47 cannot win against 46 sentences of context. It needs to be promoted to top, with structural emphasis.
How to spot it: search your prompt for the deliverable verb. If it appears once and after paragraph 2, it is buried.
3. No section headers separating task from context
Without “Background:” / “Task:” / “Output:” labels, the model has to infer structure. Inference picks the dominant theme — usually background.
How to spot it: your prompt is one flat block of prose.
4. Context is genuinely necessary but unfocused
You included 5 paragraphs because they all “felt relevant”. Half of them do not change the output. The half that does is hidden in the half that does not.
How to spot it: read each paragraph and ask “does removing this change the answer?” If no, cut it.
5. Constraints precede the deliverable
You wrote constraints, then context, then task. Constraints anchor the model to a frame, and the late deliverable cannot override it.
How to spot it: structural inversion — constraints above task.
Before you change anything
- Identify the actual deliverable in one sentence.
- Read the prompt as if you had never seen it. Where does the deliverable appear?
- Highlight any background that does not change the answer — candidates to cut.
- Decide: which 50 words of context are load-bearing?
- For repeated workflows, build a prompt template that puts deliverable first.
Information to collect
- The full prompt as sent.
- The output you got (likely a summary).
- The output you wanted.
- A pruned version of the context (only what changes the answer).
- Model and any system prompt.
Shortest path to fix
Step 1: Lead with the imperative
The first imperative sentence wins. Write it before any context:
TASK: Draft a 200-word launch email to the engineering team.
Tone: direct, no marketing speak.
Output: just the email body, no subject line.
CONTEXT: <pruned context here>The model reads “Draft a 200-word launch email” before anything else and frames everything after as input to that task.
Step 2: Use section labels
Concrete structure helps the model parse:
## Task
Draft a launch email.
## Audience
Internal engineering team, 12 people.
## Background (use only what is relevant)
- We are launching Feature X next Tuesday.
- It replaces the legacy retry logic.
- Engineering already knows about the deprecation.
## Output format
- Plain text email body.
- 150-200 words.
- No subject line.Labels turn prose into reference material.
Step 3: Prune mercilessly
For each background paragraph, ask “would the answer change if I removed this?” If no, remove it. A 600-word prompt with 200 useful words beats a 1000-word prompt with 200 useful words and 800 noise words.
Step 4: Repeat the task at the end
A single restatement at the bottom anchors recency too:
[Top: task and context]
[Bottom]
Reminder: Draft only the launch email body. No subject line.
No commentary. 150-200 words.Top + bottom hits both attention endpoints.
Step 5: Move large reference into a file or quoted block
If the background really is 2000 words, attach it as a file or fence it in a <reference> block so the model treats it as reference text, not as a request to summarize:
TASK: Draft a launch email referencing the policy below.
<reference>
... 2000 words of policy ...
</reference>
OUTPUT: Just the email body.Step 6: Test with a stranger
Send the prompt to a colleague (no context). If they cannot tell what to produce in 10 seconds, the model cannot either.
How to confirm the fix
- A stranger reads the first 3 sentences and correctly identifies the deliverable.
- The output is the deliverable, not a summary of context.
- Removing redundant context paragraphs does not change the output.
- Running the same prompt 3 times produces 3 outputs that are all the right type of thing.
If it still fails
- The context may be too large for the model to attend to the task. Cut more.
- Switch to a model with larger context and better long-context attention.
- Two-pass: first prompt summarizes the context, second prompt uses the summary + task.
- If the task itself is multi-part, split into multiple prompts.
Prevention
- Default template: TASK first, CONTEXT second, OUTPUT FORMAT last.
- Cap context: include only what changes the answer.
- For repeated workflows, save a template so you do not improvise structure each time.
- Re-read the first 3 lines of every prompt before sending; a stranger should know what to produce.
- Use section headers when context exceeds 200 words.
- Watch for “I should mention” or “for context” — these often signal cuttable paragraphs.