Reusable Claude Agent Patterns: Prompts, Tools, Context (Harnessing Claudes Intelligence)

After you have built two or three Claude agents, you start noticing that the hard parts repeat. The same context bloat, the same tool ambiguity, the same prompt that worked yesterday and confuses the model today. The teams that ship reliable agents are not smarter about each new problem — they have a library of patterns they reach for, the way a backend engineer reaches for retry-with-backoff without thinking. This post collects the reusable, code-level patterns that hold up across Claude agents in production: how to structure prompts, design tools, and manage context so your primitives compound instead of fighting you.

Prompt patterns: contracts, not conversations

The first pattern is to treat the system prompt as a contract rather than a chat opener. A production system prompt has four stable parts: the role and scope (what this agent is and is not for), the constraints (what it must never do), the tool-use policy (when to prefer a tool over answering directly), and the output contract (the exact shape of a valid final answer). Keeping these parts in a fixed order and stable wording does two things — it makes behavior predictable, and it lets you cache the prefix so repeated calls are cheaper and faster.

A second prompt pattern is the explicit decision rule. Rather than hoping the model infers when to stop or escalate, encode it: "If you cannot find a supporting fact after two tool calls, set needs_human to true and stop." Claude follows clear, testable rules far more reliably than vibes. The third pattern is the negative example sparingly used — one short illustration of a wrong output and why it is wrong does more to prevent a failure mode than three paragraphs of positive instruction.

Tool patterns: sharp names, structured returns, idempotent writes

Tools are where most agent quality is won or lost, and the patterns are concrete. Name and describe each tool for a reader who cannot see your code, because the description is the only thing the model uses to choose. Prefer a few sharp tools over many overlapping ones; when two tools could plausibly handle the same request, the model will sometimes pick wrong, and you will spend a debugging session learning which.

Always return structured, self-describing results. A tool that returns {"found": false, "reason": "no order with that id"} lets the model reason about the miss; one that throws a 500 just derails the loop. For any tool that writes — creating a ticket, sending an email, charging a card — make it idempotent by accepting a client-supplied key, so a retried call after a timeout does not double-book. This single pattern prevents a whole category of production incidents that are nearly impossible to debug after the fact.

flowchart TD
  A["Model emits tool intent"] --> B{"Args valid vs schema?"}
  B -->|No| C["Return structured validation error"]
  B -->|Yes| D{"Write operation?"}
  D -->|No| E["Execute read"]
  D -->|Yes| F["Check idempotency key"]
  F --> G["Execute once, record key"]
  E --> H["Return structured result"]
  G --> H
  C --> H

Context patterns: the relevance budget

The governing context pattern is to spend a relevance budget, not a token budget. Even with a 1M-token window, every token you add competes for the model's attention, so the question on each turn is not "can this fit?" but "does this earn its place?" Concretely: keep the stable contract in the cached prefix, keep only unresolved working memory in the middle, and retrieve documents just-in-time rather than pre-loading a knowledge base. When a sub-task resolves, compress its trail to a one-line outcome the later turns can still reference.

A closely related pattern is progressive disclosure through skills. Instead of stuffing every capability's instructions into the baseline prompt, package each as an Agent Skill — a folder of instructions and scripts Claude loads only when the task signals it is relevant. The agent stays lean by default and pulls in the refund workflow, the SQL conventions, or the brand-voice guide exactly when it needs them. This keeps the common path fast and the rare path fully equipped.

Memory and state patterns across turns

Agents that run for many turns need a deliberate memory pattern, because raw history grows without bound. The workhorse is rolling summarization: once the conversation passes a threshold, replace older turns with a structured summary that preserves decisions, open questions, and key facts while dropping the play-by-play. Keep the summary structured rather than prose — a short list of "established facts" and "open items" survives compression far better than a paragraph.

For state that must be exact, do not rely on the context at all. Hold order numbers, IDs, and amounts in your application state and pass them into context as needed, rather than trusting the model to carry a precise value across twenty turns. The pattern is: the model reasons, your code remembers. Anything that must be exactly right lives in your store and enters context as a fresh, authoritative fact.

Orchestration patterns: when to split work

For multi-step or parallelizable work, the orchestrator–subagent pattern earns its keep, with one discipline: isolate context, merge results. The orchestrator decomposes the task and gives each subagent a clean, minimal context scoped to its piece, then merges the structured outputs. Isolation is the point — a research subagent should not inherit the coding subagent's clutter. But because multi-agent runs use several times more tokens than a single agent, apply the pattern only when work genuinely parallelizes or when context isolation clearly raises quality. The matching anti-pattern is spawning subagents for sequential work that one agent could do in one focused context.

Error and recovery patterns

The last family of patterns is recovery. Wrap every tool call so a failure becomes a structured message the model can act on, not an exception that kills the loop. When a tool fails, prefer a reformulation prompt — tell Claude what went wrong and let it adjust — over a blind retry, then fall back to backoff retries, then to escalation. Cap the recovery attempts so a stubborn failure surfaces to a human instead of spinning. Combined with idempotent writes, this turns transient infrastructure hiccups from incidents into footnotes in a trace.

Frequently asked questions

What is the highest-leverage pattern to adopt first?

Structured tool returns. The moment your tools return self-describing JSON — including explicit not-found and error cases — the model's downstream decisions improve and your loop stops derailing on exceptions. It is a small change with outsized impact.

How do I keep prompts stable enough to cache but flexible enough to evolve?

Separate the stable contract (role, constraints, tool policy, output shape) from the volatile inputs (the current task and tool results). Keep the contract in a fixed, cached prefix and let only the task-specific content vary turn to turn.

When does rolling summarization hurt more than it helps?

When you summarize values that must stay exact. Summaries are for reasoning context; precise IDs, amounts, and identifiers should live in your application state and be re-injected as authoritative facts rather than compressed.

Are these patterns specific to the Agent SDK?

No. They are model-and-loop patterns that apply whether you use the Claude Agent SDK, Claude Code, or the raw Messages API. The SDK gives you convenient hooks to implement them, but the patterns themselves are portable.

Bringing these patterns to voice

CallSphere runs on exactly these patterns — contract-shaped prompts, sharp idempotent tools, and lean just-in-time context — applied to voice and chat agents that handle live calls, act on tools mid-conversation, and book work reliably. See the patterns at work at callsphere.ai.

---

Source & attribution: This is an independent, original explainer inspired by Anthropic's coverage on the Claude blog. Claude, Claude Code, Claude Cowork, Claude Opus, and the Model Context Protocol are products and trademarks of Anthropic. CallSphere is not affiliated with or endorsed by Anthropic.