How to Give an AI Agent the Right Context

Giving an AI agent the right context helps it complete a task accurately without burying important instructions in irrelevant detail. Before you start, know the outcome you want, gather the information the agent is allowed to use, and decide what it must not do.

What You’ll Need

• A specific task or outcome, such as drafting a summary or organizing a set of files.
• Relevant source material, such as a short brief, approved documents, examples, or data.
• Any constraints that affect the result: audience, format, tone, deadline, or limits on actions.
• A way to review the agent’s output and correct or approve it before anything consequential happens.

Step-by-Step

1. Define the task in one sentence. Describe the result, not a vague activity. “Draft a 200-word update for customers using the notes below” is clearer than “Help with communications.” If the task has several unrelated outcomes, split it into separate requests.
2. State what a good result looks like. Specify the intended reader, purpose, format, and level of detail. For example: “Use three short sections, plain language, and only information supported by the supplied notes.” Include a length range only when it matters. Avoid instructions that conflict, such as asking for both a comprehensive report and a very short answer.
3. Provide only material that changes the answer. Include facts, definitions, decisions, and examples the agent needs. Leave out background that is merely interesting. A useful test is: would removing this item change the result or help prevent a mistake? If not, omit it or make it available separately rather than placing it in the main prompt.
4. Separate instructions from source material. Label the parts clearly, for example with “Task,” “Constraints,” and “Reference notes.” Say whether the agent should treat supplied documents as evidence, follow a style example, or simply summarize them. This reduces confusion when a document contains text that looks like an instruction.
5. Set boundaries and authority. Explain what the agent may do and what requires your approval. For instance, it may draft a message but must not send it, or it may suggest file changes but must not apply them. State which sources to rely on and how to handle missing, uncertain, or conflicting information. Ask it to flag gaps instead of guessing.
6. Give a small, relevant example if the format is hard to describe. A short example can demonstrate tone, structure, or the difference between an acceptable and unacceptable result. Use an example that resembles the task, and identify which features to copy. Do not provide a large collection of examples that distracts from the actual request.
7. Keep long-term preferences separate from task-specific facts. If the agent supports saved instructions or project notes, use them for stable preferences that apply repeatedly. Put temporary details—such as this task’s audience, deadline, or source data—in the current request. Check saved context periodically so outdated preferences do not distort new work.
8. Ask for a check before finalizing. Request that the agent compare its draft with the stated requirements, identify unsupported claims, and list unresolved questions. For tasks with multiple requirements, provide a short checklist. Review the output yourself, especially facts, calculations, citations, and any action that affects people or systems.
9. Refine based on the result. If the response misses the mark, identify the specific failure and change the relevant instruction or source. Say “Use only the two approved documents; the third paragraph is unsupported,” rather than adding a broad pile of new background. Make one or two focused changes at a time so you can see what helped.

Common Mistakes

• Sending everything “just in case.” Large, mixed collections make relevant details harder to distinguish. Select the necessary sections, or summarize them and point to the exact source passages.
• Leaving the goal implicit. The agent cannot reliably infer whether you want a critique, a rewrite, a decision, or a list of options. Name the requested deliverable.
• Mixing requirements with preferences. Mark true must-haves separately from optional guidance. For example, “Do not include personal data” is a constraint; “Prefer a friendly tone” is a preference.
• Using vague references. Words like “this,” “latest,” or “the usual version” may be unclear. Name the document, date, audience, or version where it matters.
• Assuming the agent remembers or verifies everything. Context may not carry between conversations, and generated answers can still be wrong. Repeat essential details and provide authoritative sources when accuracy depends on them.
• Giving conflicting instructions. If a template says one thing and the prompt says another, tell the agent which takes priority. Remove outdated guidance when possible.

Tips

• Put the most important goal and constraints near the beginning, then add supporting detail.
• Use compact headings and bullets so requirements are easy to scan.
• Ask for a particular uncertainty behavior: “If a required fact is missing, ask me; do not fill it in.”
• For a multi-stage task, provide context in stages: first agree on the plan, then supply the material needed for each step.
• Use a concise reusable brief for repeated work, but review it whenever the task, audience, or source of truth changes.
• Remove confidential or personal information that is not necessary, and follow your organization’s rules for sharing data with the tool.

Troubleshooting

The answer is generic. Add the intended audience, purpose, and one or two concrete examples of what a useful answer should include. Provide the source material if the task depends on facts it cannot know.

It ignores a key detail. Move that requirement into a clearly labeled constraint near the top. If there are many requirements, ask the agent to repeat them as a checklist before it drafts, then check the result against each item.

It invents or misstates facts. Tell it to use only specified sources, cite or identify the supporting passage when possible, and mark anything the sources do not establish. Verify important claims against the original material; a confident tone is not proof.

It gets worse after you add more context. Remove repeated, stale, or unrelated material. Keep the best source and summarize the rest into a short set of facts. Make the goal and priority rules explicit.

It takes an action you did not intend. Pause the workflow and clarify permissions before continuing. Require a confirmation step for sending, publishing, editing, deleting, or other consequential actions. If a change has already occurred, use the relevant system’s normal review or recovery process.

Frequently Asked Questions

How much context should I give an AI agent?
Give enough to define the task, provide necessary evidence, and state constraints. There is no universal ideal length; relevance matters more than volume.

Should I put instructions before or after the source material?
Put the goal and critical constraints first, then label source material clearly. For long inputs, restate the requested outcome at the end if needed.

Can I give an agent a whole document or folder?
You can when it is relevant and the tool supports it, but identify the specific sections or files to use. A focused excerpt is often easier to work with than a large, mixed collection.

What should I do when the agent needs information I have not supplied?
Tell it to ask a question or list the missing information rather than assume. Provide the missing facts only after checking that they are accurate and appropriate to share.

Does saved memory replace instructions in a new task?
No. Saved preferences may help with recurring work, but task-specific goals, constraints, and source material should be stated in the request. Check that stored context is still current.

How can I tell whether the context is enough?
Check whether the agent can identify the desired output, the evidence it should use, the constraints it must follow, and what to do if information is missing. Review its result against those points before relying on it.

Related

• On-Device LLMs: How Hybrid AI Balances Privacy and Power
• Deno 2.5, Pragmatic Turn, Bytes and 423 Explained
• Convert AVI to PSP: Format, Settings, and Transfer