Pattern Catalog

Start with general background, progressively add specific details. Give the model altitude before asking it to land. Mirrors how experts brief each other; context first, task second.

Include the smallest set of high-signal tokens that helps the model do the task. More context usually means weaker attention.

When conversations grow long, summarize what matters and start fresh. Context quality degrades well before hitting advertised limits.

Persist important context to external storage: scratchpads, memory files, knowledge bases. The context window is working memory, not long-term memory.

Retrieval gets information into context. Grounding makes the model actually use it. Without explicit anchoring instructions, the model will often ignore what you retrieved and fall back to whatever it absorbed during training.

Front-load all source reads into one turn so every subsequent turn works from cache.

Give sub-agents their own focused contexts instead of sharing one massive window. Anthropic's multi-agent system uses 15x more tokens total but gets better results, because each agent sees only what it needs.

Let agents spawn child agents with scoped sub-contexts. Instead of stuffing everything into one window, the parent splits work, delegates with focused context, and aggregates results.

Start with an index of available context. Let the model pull in details on demand instead of loading everything upfront.

A JSON schema tells the model what to think about, in what order, and with what vocabulary. Define the structure and the model's reasoning follows.

Reuse computed context across requests to reduce costs and latency. Structure prompts so the stable prefix gets cached and only the variable part changes.

Place critical information at the start and end of context. Models over-attend to the beginning and end of their context window, a phenomenon called 'lost in the middle.' Work with this bias instead of against it.

Weight recent context higher and systematically age out old information. Old messages can stay available, but they should compete less with the current task.

Tool definitions are context. The description tells the model when to use a tool and how. Most descriptions only say what the tool does; the ones that work also say when to use it and when not to.

Choose examples that resemble the current input, even when the easiest examples to find look cleaner. The wrong examples teach the model the wrong behavior.

Treat the context window as a finite resource with planned allocations. Decide upfront how many tokens each section gets, then enforce it.

Defining a role in the system prompt does more than set a tone. It activates a vocabulary, constrains scope, and steers which heuristics the model applies. The specificity of the role determines how much of that steering actually lands.

Images consume tokens aggressively and at unpredictable rates. Choose the right modality for each piece of context (raw image, text description, or structured extraction) before the model sees it.

"Don't do X" is weaker than it looks. Negative instructions activate attention on the prohibited thing and leave the model without a path forward. Reserve them for hard stops; use positive framing everywhere else.

When one agent passes work to another, most of the context gets lost. The handoff boundary is where multi-agent systems silently degrade, because nobody designed what travels with the task.

A hallucination in the context window becomes ground truth for every subsequent turn. The model generated it, so it trusts it, and the error compounds silently until the output is confidently wrong about something that was never true.

Retrieval isn't just search. Every retrieval decision is a context engineering decision: what to retrieve, how much, in what order, and what to leave out. The vector store returns candidates; you decide what earns a place in the window.

Not all context is created equal. System instructions, user messages, retrieved documents, and tool outputs compete for the model's attention, and without explicit priority signals the model resolves conflicts unpredictably.

Maintain structured working state inside the context window: a running plan, a list of findings, a set of decisions made so far. Without an explicit scratchpad, the model reconstructs its state from raw conversation history on every turn, and gets worse at it as the conversation grows.

Split context retrieval into a focused agent that returns exact evidence. The main agent should receive selected files, ranges, and facts after the search noise has been discarded.

Treat every summary as a risky rewrite. Validate compressed context against the next task before you trust it.

Select memories by their measured effect on the answer. A memory belongs in context only if it improves the next step.

Clean unsafe or adversarial state before it enters memory, summaries, or handoffs. Sanitizing only the final summary is too late.

Persist the agent's evidence trail alongside the artifact it changed. Future agents need the reasoning path and the final diff together.