Answer
How do you give an AI enough context to do a job properly?
Supply the material, the standard of a good result, and the constraints. Most failures are missing constraints, not missing information.
Give it the actual source material, a description of what a finished result looks like, and the constraints that make a wrong answer wrong. Most disappointing output comes from missing constraints rather than from missing information.
There are three kinds of context and they do different work. Material is the documents, records and examples the answer must be built from. Standard is what a finished, acceptable result looks like. Constraint is what rules an answer out. People supply material generously, standard occasionally, and constraint almost never — and constraint is the one that separates useful output from plausible output.
Material first, because it is the largest single improvement available and it is free. A question answered from a document you supplied is a different operation from a question answered from general knowledge. Paste the policy, the specification, the last three examples of the thing you want, the actual customer record. The instinct to describe the material rather than supply it is the most expensive habit in ordinary AI use.
Standard is the part that feels unnecessary and is not. If you cannot state what a good result looks like, the system will produce the median of everything that has ever been called that kind of document, which is exactly the flat, competent, useless output people complain about. The efficient form of a standard is an example: one previous output you were happy with does more work than a paragraph describing your preferences, because it carries structure, length, register and level of detail at once.
Constraint is what is being asked for when someone says the answer was technically right and completely wrong. It includes the things that must not appear, the assumptions that are not available, the audience, and the decisions that have already been taken and are not up for reconsideration. A brief that omits the decisions already taken invites the system to relitigate them helpfully, and the output then has to be argued with rather than used.
There is a fourth thing that is not context and is frequently confused with it: the task. Context describes the situation; the task says what to produce. A common failure is a long, rich brief that never states the deliverable, which produces an essay about the situation. Keeping the two visually separate — situation here, deliverable there — costs nothing and fixes a large share of vague output.
More context is not monotonically better, and this is where the intuition breaks down. Beyond the point where the material is present, adding volume dilutes rather than informs: the specific instruction sits among many pages of background, and the system weights it accordingly. The discipline is to supply what bears on this job and leave out what merely relates to the subject, which is a harder editorial decision than dumping the folder in.
A model asked for a good answer will produce an average one, because average is what good means without a stated standard.
Siddharth Sharma, Context Theory
Related questions
How much material is too much?
The signal is not a token count, it is when you can no longer say what each part of the context is for. If a document is in there because it is related rather than because a specific decision in the task depends on it, it is diluting the parts that do. The practical fix is to supply the relevant extracts rather than the whole file, which also makes it obvious when the extract does not actually contain the answer.
Is it better to give context in one message or build it up?
One assembled brief is more reliable for a defined job, because everything is present when the first decision is made. Building up works for exploration, where you do not yet know what matters, but it leaves the early reasoning based on a partial picture and that reasoning tends to persist. If a conversation has turned into a job, it is usually worth restating the brief once, in full, before asking for the deliverable.
METHOD
Every figure below carries its source and the date it was verified. Nothing on this page is asserted.
The numbers on this page.
| What | Value | Specific to |
|---|---|---|
| Visibility lift in AI-generated answers from GEO methods | up to 40% | Category-wide |
| AI-cited sources that also rank in the Google organic top 10 | 10% | Category-wide |
Aggarwal et al., "GEO: Generative Engine Optimization", Princeton / Georgia Tech / IIT Delhi / Allen Institute for AI — KDD 2024 · GEO-bench · 10,000 queries across 8 domains · verified
2026 generative engine citation study · fewer than · verified
What is specific to this page.
| Kind | Claim | Check it against |
|---|---|---|
| Workflow | Context divides into material, standard and constraint, and the constraint class — what rules an answer out, what has already been decided, what must not appear — is the one most often omitted and the one that separates useful output from merely plausible output. | Auditing a set of unsatisfactory outputs against the brief that produced them and classifying which of the three was absent. |
| Software | A single example of an acceptable previous output conveys structure, length, register and depth simultaneously, which is why it outperforms a written description of the same preferences as a way of stating the standard. | Running the same task with a described standard and with one worked example, and comparing which output needs less revision. |
| Buying behaviour | Beyond the point where the necessary material is present, additional context dilutes rather than informs, because a specific instruction carries less relative weight when embedded among a large volume of merely related background. | Repeating a task with the relevant extract alone and with the whole source folder, and comparing adherence to the specific instruction. |
| Response | Situation and deliverable are distinct and a brief that supplies rich context without naming the artefact to produce reliably returns discussion of the situation rather than the artefact. | Checking any unsatisfactory brief for a sentence that names the thing to be produced. |
Each row would be wrong on another industry's page. Where a sourced figure exists it is in the table above instead; these are the constraints that shape the work and do not happen to be numbers.
Start with the measurement.
Reading about a benchmark is not the same as knowing your own number. The audit produces yours, measured rather than estimated.
$497 · delivered in 5 business days · credited against month one