Anthropic: Effective context engineering
Guidance on selecting and maintaining instructions, data, history, tools, and agent state.
← Back to Compute College
Compute College
What belongs in context?
Select the smallest set of relevant, authoritative, recent, and complete information needed for the next AI decision.
Plain-English definition
Information belongs in context when it is needed for the current decision, trustworthy for that decision, current enough to use, and complete enough to avoid a misleading answer. Relevance alone is not enough: an old policy can be relevant but wrong, and a current fragment can be authoritative but incomplete without its surrounding definition.
Memory trick: Relevant is not enough. Context should be relevant, right, recent, and sufficient.
Why it matters
Context selection affects answer quality and recurring workload cost. Sending everything increases tokens and makes important details harder to find; sending too little creates unsupported answers and retries. A clear inclusion rule also helps the application distinguish instructions, evidence, memory, and untrusted user content.
- Authority determines which source wins when documents disagree.
- Recency matters for policies, prices, permissions, and changing product state.
- Completeness prevents a short but misleading passage from being treated as the entire answer.
Simple example
A refund agent may need the current refund policy, the order record, delivery status, and the customer’s current request. It does not need every customer order, the company’s entire policy library, a superseded policy version, or a completed tool result from an earlier task.
- Include the source needed for the decision, not the whole repository.
- Keep policy version and effective date alongside the retrieved passage.
- Keep user-provided claims distinguishable from authoritative records.
Example figures are illustrative calculations, not current quoted market prices.
Current example
Context selection reference
Anthropic’s context-engineering guidance frames context as a curated set of tokens and information, not a document dump. That framing supports explicit decisions about what is relevant, authoritative, and useful at each step.
Source discipline: the reference explains the concept or method; it is not a substitute for measuring a production workload under its actual provider, model, and data conditions.
Common mistake
“Related” does not automatically mean “belongs in context.” A document can mention the right topic while using an old definition, an unsupported assumption, or a different customer or region. Relevance needs an authority and freshness check.
Practical takeaway
What you can do with this
For one AI feature, create an inclusion checklist with relevance, authority, recency, completeness, and sensitivity. Apply it to ten retrieved items and mark why each item entered or stayed out of context.
- Knowledge teams: attach source, owner, version, and effective date metadata.
- Builders: expose retrieval reasons and citations to evaluators.
- Operators: watch stale-source and conflict rates as the corpus changes.
Decision check: can the system explain why each context item is necessary and which source should win if two items disagree?
Compute College track
Prompt & Context Engineering
Step 11 of 18: What belongs in context