Context Is All You Need Prompting
Treat the model as a brilliant stranger with zero context, and supply what a colleague would already know.
- Difficulty
- Easy
- Time to result
- ~days to results
- Steps
- 5
- Confidence
- 95%
Logan Kilpatrick's core theory of prompt engineering: models give generic answers not because they are dumb, but because they are 'the most eager person in the world' answering a question with no context about you, your goals, or your standards. The fix is not magic words — it is systematically loading the context a knowledgeable human colleague would already possess. He models prompting as the same act of effective communication you'd use to get value out of a person.
Origin
Logan Kilpatrick's own working theory as head of developer relations at OpenAI, articulated on Lenny's Podcast (Feb 2024). The high-fidelity-rewrite step is modelled on how DALL·E internally expands a short image prompt. The interview-prep prompt example was contributed by Lenny Rachitsky, who credits Tim Ferriss for the underlying idea.
Core principles
- 01Crap in, crap out — a basic question yields a basic answer, exactly as with humans.
- 02The model appears smart, so we wrongly assume it has context. It has none.
- 03The model will never tell you it lacks context — it wants to answer regardless.
- 04Prompt engineering is a human communication skill, not a technical incantation.
- 05Tone and framing (a smiley face, 'take a break') give small real gains because the training corpus is human-to-human messages.
How to run it
- 1
Write the naive prompt, then interrogate it
Draft the prompt as you naturally would, then ask: what would a smart colleague need to know before they could answer this well? Everything on that list is missing context.
Pro tip Assume the model has human-level intelligence and literally zero knowledge of you, your goals, or your audience.
Watch out The model will happily return a confident, generic answer rather than telling you it lacks information. Silence is not a signal that your prompt was sufficient.
- 2
Test whether the subject exists in the training data
Ask whether the person, product, or topic has enough public footprint for the model to have real knowledge. Famous subjects work with thin prompts; niche subjects need you to supply the corpus yourself.
Pro tip Rule of thumb from Kilpatrick: a prompt about Tom Cruise works fine; the same prompt about a lesser-known operator returns filler.
- 3
Feed the corpus, don't describe it
Rather than naming the subject, paste or link the raw material — blogs, tweets, docs, transcripts — and instruct the model to read it and surface the interesting angles.
Pro tip Use browsing or file upload so the model works from primary source material rather than its own vague recollections.
- 4
Do the high-fidelity rewrite the model would do for images
Expand 'write me a blog post about AI' into an explicit brief: the trade-offs to cover, the use cases to include, the papers to reference, the audience, the tone. This mirrors how DALL·E silently expands a two-word image prompt into a detailed one.
Pro tip Ask the model to first expand your prompt into a high-fidelity version, review and edit it, then run it. You become the editor rather than the typist.
- 5
Layer the small human touches last
Add the low-order tweaks — a smiley face, telling the model to take a break before answering — but treat them as a 1-2% effect, not a strategy.
Pro tip These marginal gains only become discernible on long generations; on a few-sentence answer they are noise.
Watch out Do not substitute cosmetic tricks for context. They cannot rescue a context-starved prompt.
In the wild
Lenny asked ChatGPT for ten questions to ask Logan Kilpatrick in the style of Tyler Cowen. The questions came back 'fine' but not exceptional. Kilpatrick diagnosed the problem instantly: there isn't enough about him on the internet for the model to have real context, so it padded. The fix is to hand it his blog and his Twitter and instruct it to mine his actual perspectives.
→ The same prompt shape works well for a heavily-documented subject and fails for a niche one — the variable is context, not prompt craft.
When a user asks DALL·E for 'a picture of a turtle', the system silently rewrites the request into a high-fidelity description — a turtle with a shell, a green background, lily pads in the water — because the model was trained on richly-detailed examples.
→ Kilpatrick uses this as the template for text prompting: the fidelity gap you feel in the output is the fidelity gap you left in the input.
Common mistakes
Assuming the model knows who you are
Because the system sounds intelligent, users project omniscience onto it. It has never met you, does not know your goals, and will not ask.
Chasing prompt hacks instead of loading context
Smiley faces and 'take a break' produce roughly 1-2% gains. Context produces order-of-magnitude gains. Optimising the former while ignoring the latter is the most common failure.
Being lazy about typing what you actually mean
Kilpatrick names human laziness as the root cause: we don't want to type what we mean, so we under-specify and then blame the output.
Is it for you?
Best for
Knowledge workers, PMs, and founders who use ChatGPT daily and keep getting technically-correct but generic output they have to rewrite anyway
Not ideal for
Highly-templated, repetitive tasks where a fixed system prompt already encodes the context once and for all
From the transcript
“prompt engineering is a very human thing like when we want to get some value out of a human we do this prompt engineering we…”
“the reason that you get super generic responses sometimes is because people forget they need to put that context in the model”
“context is the only thing that matters”
“and again giving the model enough context to answer to answer the question”
From the episode
Inside OpenAI
Logan Kilpatrick (head of developer relations)