Reusable Patterns for Building Claude Cowork Agents

The first Claude Cowork plugin a team ships usually works. The tenth one is where the trouble starts — skills overlap, two of them try to load on the same request, a tool returns data in a shape nothing downstream expects, and nobody can tell why the agent suddenly picks the wrong procedure. The fix is not more prompting; it's structure. The teams that scale Cowork cleanly treat skills, tools, and context as software components with contracts between them. This post lays out the reusable patterns that make that possible, with the actual shapes you'll write.

Key takeaways

• Give every skill a single responsibility and a non-overlapping trigger description so the model never has to guess between two.
• Treat tool outputs as typed contracts — pin the shape so skills can rely on it instead of re-parsing prose.
• Layer context: stable instructions up top, task-specific detail loaded on demand, transient data last.
• Make tools idempotent and name them by intent so the model picks correctly from the description alone.
• Compose small skills via explicit handoffs rather than one giant skill that tries to do everything.

Pattern 1: One skill, one responsibility

The most common scaling failure is the kitchen-sink skill — a single folder that triages, summarizes, drafts replies, and updates records. It works until two of its jobs need to behave differently, and then every edit risks the others. The reusable pattern is one skill per coherent responsibility, each with a description that does not overlap any other. When responsibilities are clean, the model's loading decision is unambiguous: only one description matches.

Concretely, split "handle tickets" into ticket-triage, ticket-summary, and ticket-reply-draft. Each is short, testable, and editable in isolation. The cost is more skill folders; the benefit is that a change to the reply drafter can't break triage, and the model never wavers between two similar descriptions.

Pattern 2: Tool outputs as typed contracts

When a tool returns free-form text, every skill that consumes it has to re-interpret prose, and small wording changes break downstream behavior silently. The durable pattern is to make tools return a fixed, documented shape and to write skills against that shape. Pin the contract once and reuse it everywhere.

// Tool contract: search_tickets always returns this shape
{
  "tickets": [
    {
      "id": "OPS-4821",
      "severity": "high",         // one of: low|medium|high|critical
      "sla_breached": false,
      "summary": "string",
      "updated_at": "2026-06-06T14:00:00Z"
    }
  ],
  "total": 1
}

Now ticket-triage can reliably group by severity and flag sla_breached without parsing sentences. When you later add a field, existing skills keep working because the contract only grew. This is the same discipline as a stable API response — the model is just another consumer of it.

flowchart TD
  A["User request"] --> B["Match exactly one skill description"]
  B --> C["Skill calls intent-named tool"]
  C --> D["Tool returns typed contract"]
  D --> E{"More work needed?"}
  E -->|Same domain| C
  E -->|Other domain| F["Handoff to sibling skill"]
  E -->|Done| G["Compose answer from typed fields"]

Pattern 3: Layered context, loaded by relevance

Context is a budget, and the pattern that keeps agents sharp is layering it by stability and relevance. Stable, always-true instructions (tone, hard constraints, what the agent must never do) belong in the base context. Task-specific procedures belong in skills loaded only when relevant. Transient data — the actual ticket list — enters last and lowest. The closer something is to "true for every request," the higher and more permanent it should sit.

This layering is what makes the dynamic skill-loading design pay off. If you instead pour every procedure into the base context, the model reasons across a wall of mostly-irrelevant instructions on every turn, and quality drops. Keep the base lean and let relevance pull in the rest.

Pattern 4: Name tools by intent, make them idempotent

The model chooses tools largely from their names and descriptions, so name them for what the user wants, not for the internal endpoint. draft_ticket_reply reads as intent; post_v2_comment reads as plumbing and invites wrong calls. Pair good names with idempotency: a tool that's safe to call twice protects you when the agent retries after an ambiguous result, which it sometimes will.

// Good: intent-named, idempotent via a client key
draft_ticket_reply({
  ticket_id: "OPS-4821",
  tone: "concise",
  idempotency_key: "OPS-4821-reply-2026-06-06"
})

The idempotency key means a retried draft doesn't create two drafts. This single habit eliminates a whole class of duplicate-action bugs that otherwise surface only under load, when retries are most likely.

Pattern 5: Compose with explicit handoffs

Small skills are only valuable if they combine cleanly. The pattern is explicit handoff: a skill that finishes its job and names the next skill, rather than trying to absorb the next job. ticket-triage ends by saying "for the top at-risk ticket, hand off to ticket-reply-draft." The orchestrator then loads the sibling skill with a focused brief. This keeps each skill single-purpose while still supporting multi-step work, and it makes the chain visible and debuggable.

Common pitfalls

• Overlapping skill descriptions. If two descriptions could both match a request, the model's choice becomes unstable. Make each description disjoint, ideally keyed on a distinct verb-plus-object.
• Skills that parse prose. Consuming free-text tool output couples your skill to exact wording. Return typed contracts and read fields, not sentences.
• Context dumping. Putting every procedure in the base context defeats dynamic loading and dilutes reasoning. Reserve the base for what's true on every request.
• Endpoint-named tools. Names like get_v3_data make the model guess. Name by intent so selection is obvious from the description.
• Non-idempotent write tools. Without an idempotency key, a single agent retry can double-post. Add the key to every state-changing tool.

Apply these patterns in 5 steps

1. Split any kitchen-sink skill into single-responsibility skills with disjoint descriptions.
2. Document a fixed output contract for each tool and rewrite skills to read its fields.
3. Move always-true rules into base context and leave procedures in on-demand skills.
4. Rename tools to intent-based names and add idempotency keys to every write.
5. Replace multi-job skills with small skills connected by explicit handoffs.

Kitchen-sink skill vs. composed skills

Quality	One big skill	Composed small skills
Trigger clarity	Ambiguous	One match per request
Editability	Risky — changes ripple	Isolated
Testability	Hard to cover	Each tested alone
Reuse	Low	High via handoffs

Frequently asked questions

How do I stop two skills from both trying to load?

Give each skill a single responsibility and a description that doesn't overlap any other — ideally a distinct verb-and-object. Because Cowork loads skills by matching the description to the task, disjoint descriptions make the choice deterministic.

Why return typed tool outputs instead of text?

A fixed, documented output shape lets skills read specific fields instead of re-parsing prose, so wording changes don't silently break downstream behavior. It's the same reason you'd version an API response rather than return free text.

What belongs in base context versus a skill?

Put rules that are true for every request — tone, hard limits, must-never-do constraints — in base context, and put task-specific procedures in skills loaded on demand. The more universal a rule, the higher and more permanent it should sit.

Why add idempotency keys to tools?

Agents sometimes retry after an ambiguous result. An idempotency key makes a repeated call safe, so a retried write produces one action instead of two — eliminating duplicate-action bugs that mostly appear under load.

Bringing agentic AI to your phone lines

These composition patterns translate directly to conversation. CallSphere structures its voice and chat agents the same way — small, single-purpose skills calling typed tools — so they answer every call, act mid-conversation, and book work 24/7. See it 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.