Context Loading (Additional Information)
Front-load all relevant task information — and put it at the top for caching
- Difficulty
- Easy
- Time to result
- ~days to results
- Steps
- 3
- Confidence
- 93%
Give the model as much relevant information about the task as possible — your bio, work history, a company profile, definitions of domain terms. Schulhoff prefers the term 'additional information' over the overloaded 'context.' Placement matters: put it at the beginning of the prompt so it can be cached (cheaper repeat calls) and so the model doesn't forget its task or latch onto a stray question buried at the end.
Origin
A core prompt component identified in The Prompt Report by Sander Schulhoff. He deliberately renames 'context' to 'additional information' to avoid confusion with 'context window.'
Core principles
- 01More relevant information almost always helps — in conversational mode, go 'hog wild'
- 02In product mode, latency and per-token cost force you to trim to what's truly needed
- 03Placing additional information at the top enables provider-side caching (cheaper, faster repeat calls)
- 04Trailing long context can make the model forget its original task or hijack a question from within the context
How to run it
- 1
Gather everything relevant to the task
Assemble biographies, work history, company profiles, or definitions of any domain terms the model may not know.
Pro tip If the model doesn't know a niche term, paste in research explaining it before asking it to use that term.
- 2
Place it at the beginning of the prompt
Put the additional information block first, before your directive, so providers can cache it and the model keeps your actual task in focus.
Pro tip Cached leading context makes subsequent calls cheaper because the provider stores the context and its embeddings.
Watch out Additional information at the end of a very long prompt can cause the model to forget the task or answer a stray question found inside it.
- 3
Right-size for the setting
In conversation, dump generously. In a product with millions of API calls, identify the minimum information needed to control cost and latency.
Pro tip Structured XML wrapping is optional here — Schulhoff usually just tosses the information in without special formatting.
In the wild
Classifying whether Reddit posts showed 'entrapment' (a suicide-risk signal), GPT-4 didn't know the term, so Schulhoff pasted in research plus the professor's original email describing the problem. When the email was later removed to protect personal information, accuracy 'dropped off a cliff'; even anonymizing the names in it degraded performance.
→ Including the rich additional-information context was essential to a performant prompt; removing it collapsed accuracy.
Common mistakes
Burying context at the end of a long prompt
The model may lose track of the original directive or answer an unrelated question embedded in the trailing context.
Over-stuffing a production prompt
In product mode, dumping everything spikes token cost and latency across millions of calls.
Is it for you?
Best for
Both conversational users needing better answers and product teams tuning a critical prompt (with cost/latency discipline)
Not ideal for
High-volume production prompts where indiscriminate context blows up cost and latency
From the transcript
“I tried to call this additional information since context is a really overloaded term”
“One, it can get cached. So subsequent calls to the LM with that same context at the top of the prompt uh are cheaper”
“the the model can like forget what its original task was and might pick up some question in the additional information”
“The performance also dropped off a cliff”
From the episode
AI prompt engineering in 2025: What works and what doesn’t
Sander Schulhoff (Learn Prompting, HackAPrompt)