AI Guides › Prompting
The Difference Between Instructions And Context
By Nigel Guy · 2 min read
Most prompts are written as one undifferentiated block: background facts,
what already exists, and what you actually want done, all run together in
the order they occurred to you. It reads fine to you, because you already
know which sentence is a fact and which is an order. The model has to guess,
and it guesses wrong often enough that the answer either ignores something
important or treats a stray fact as an instruction.
The rule: context is what's true, instructions are what you want done, and
a prompt that doesn't separate the two will occasionally get them
backwards.
The mechanism: the Instructions/Context split
- Write the context block first, and label it as context. Background,
existing material, constraints that already exist independent of this
request — things that are true whether or not the model does anything
with them.
- Write the instructions block second, and label it as instructions.
Imperative, action-first sentences: what to produce, in what form, using
which parts of the context.
- Check for leakage. Read back through the context block looking for
anything phrased as a command, and through the instructions block looking
for anything that's actually just a fact. Move each to where it belongs.
- Put instructions last, closest to the actual request. Whatever
immediately precedes "now do X" tends to get weighted as most relevant —
put your real ask there, not three paragraphs of background.
A minimal version of this is two labelled headings — "Context:" and
"Instructions:" — with nothing more clever than that. The labelling is doing
almost all of the work.
What to skip
Skip labelling for genuinely short prompts — a one-line request with one
line of background doesn't need headings to stay unambiguous. Skip
re-typing context the model already has from earlier in the same
conversation; repeating it usually just adds noise, and if you're not sure
it's still in view, it's faster to ask than to restate everything.
Guardrails
- This fixes a specific failure mode — instructions and context blurring
together — not vague or badly worded instructions themselves. Pair it with
a clear task statement (see the three-part structure), not instead of one.
- Separating the blocks doesn't make the context accurate. If the
background you're feeding in is wrong, labelling it clearly just makes the
model wrong in a clearer way.
- On a long or messy prompt, do a final read of just the instructions block
on its own — if it doesn't stand up as a clear, complete request without
the context around it, it's still missing something.
All 751 AI guides · JulieMango plans from £17/mo