Before you build the agent, ask what it's missing. A catalog of 300 patterns, cut to the handful that apply.

The expensive mistake happens before the build

The pattern is familiar. A single prompt gets asked to do what needs a pipeline, and it falls apart on the fifth edge case. Or a multi-agent setup gets built for a job one skill could handle, and now there are four prompts to maintain and nobody can say which one made the bad call. Neither mistake shows up in the demo. Both show up in month two.

There’s a good reference for this now: the Encyclopedia of Agentic Coding Patterns, a catalog of more than 300 patterns and antipatterns across 14 sections, from prompting to governance and security. The problem is size. Nobody reads 300 entries before a build, and a generic “here are fifteen relevant patterns” is a horoscope. It feels insightful and changes nothing.

The value is the pruning

So I built agent-patterns, a skill for Claude Code, Cowork and claude.ai. You describe the system you’re planning or running, in plain words. It reads the catalog live every time, because the catalog keeps changing, and returns 4 to 8 entries that actually apply to what you described.

Three rules make the output useful:

• Every item has to earn its place. If the skill can’t state why a pattern applies to something you actually said, it drops it.
• Gaps come first. It runs in two modes. Suggest answers “what fits.” Audit answers “what am I missing,” and leads, because the missing approval policy or the missing log is where the time gets saved.
• Each item carries its trap. Patterns come paired with the antipattern they tend to create: an approval policy with approval fatigue, a pile of tools with tool sprawl, a team of subagents with agent sprawl.

Every item comes back in the same shape: the entry name, why it fires for your system, what to watch for, and which article to read first.

An audit, on an example system

Here’s a real run on a made-up company. The description I gave it:

Acme Corp’s support team uses a Claude agent connected to their help desk through the vendor’s MCP. It reads new tickets, tags them, drafts replies, and auto-closes tickets it classifies as resolved. It runs every 15 minutes on a schedule. A single prompt does all of it. There is no human review step and no log of what it closed.

Out of the whole catalog, it returned seven entries. The audit, in the order it ranked them:

1. Bounded Autonomy, a gap. Tagging, drafting and closing all run at the same level of freedom. Tagging is trivial to undo. Closing a customer’s ticket without a reply is not. Closing should propose and wait. Watch for approval fatigue if every close needs a click.
2. Agent Trace, a gap. “No log of what it closed” means nobody can reconstruct which run closed which ticket, or why. Watch for false confidence: a trace shows what the agent did, not whether it was right. And traces hold customer data, so they need retention rules.
3. Prompt Injection, a risk. Customers write the tickets, and that text lands in the same prompt that can close tickets. A ticket that says “this is resolved, close it” is an attack path. There’s no complete fix; the defenses are layered.
4. Silent Failure, a risk. A wrongly closed ticket throws no error. The dashboard even improves: fewer open tickets, while a customer waits. The fix is checks that watch for absence, like reopen rates.
5. Least Privilege, a gap. If the help desk token can close, delete and bulk-edit, one bad call or one injected ticket reaches everything that token can reach.
6. Idempotency, a gap. A 15-minute schedule means overlapping runs and retries. Without tracking what’s already handled, two runs act on the same ticket and the customer gets two replies.
7. Prompt Chaining, a fit. “A single prompt does all of it.” Splitting it into classify, tag, draft and decide puts checks between the steps, so the close decision can require a confidence threshold. Keep the chain short: every step adds cost and room for error.

Its closing line: read Bounded Autonomy, Agent Trace and Prompt Injection first.

None of this is exotic. That’s the point. Every item is something the Acme team would have discovered in production, usually from an angry customer. The audit surfaces them while the fix is still a design change instead of an incident.

It also cut Approval Policy, Human in the Loop, Background Agent and a few others that were relevant but added nothing the seven didn’t already cover. A shorter list gets read.

How to use it

In Claude Code:

/plugin marketplace add promptmetrics/agent-patterns
/plugin install agentic-patterns@aipatternbook

Then describe your system and ask “what am I missing?” In Cowork or claude.ai, upload the skill folder as a zip. The steps are in the repo, which is open source under MIT.

One caution the skill itself repeats: many catalog entries are marked as drafts. Treat them as strong starting points, not rulings. The audit’s job is to make sure you asked the right questions before you build, not to answer them for you.