Claude Prompts for Building AI Agents
Updated November 2026 · 15 prompts
Quick answer
Use Claude to design the parts that decide whether an agent works: the system prompt, the tool descriptions, an eval set, and a review of failed runs. The 10 prompts below cover each step, written for Claude's habit of following explained instructions better than rules in capitals. Five more keep long agent-debugging chats within your usage limit.
ClaudeKit adds this pack to Claude: pick "AI agent builders" when it opens, then insert any prompt in one tap. It also shows your Claude usage limit as you work. Free, no account.
Add to Chrome, freeMost agent failures aren't model failures. They come from a tool description that doesn't say when to use the tool, a system prompt that never says when to stop, or no test set to tell whether a change helped. These prompts turn Claude into a reviewer for each of those pieces.
Agent work is also hard on usage limits: transcripts are long, and pasting three of them into one chat fills it fast. The Limit Saver prompts at the end help you summarize a run and start fresh.
How to get the most from these prompts on Claude
- Paste real transcripts and real tool schemas. Claude can only find the first wrong step if it can see it.
- Explain why an instruction matters instead of writing it in capitals. Claude follows reasons better than emphasis, and heavy emphasis can make an agent over-apply a rule.
- Keep one chat per agent problem, and summarize a long run before you start the next.
- Re-run your eval set after every prompt or tool change, so you know whether it helped.
The AI agent builders prompts
Copy a prompt, fill in the parts in [square brackets], and paste it into Claude.
Write a system prompt
Write a system prompt for an agent that [does the job]. Cover its role in one sentence, what good output looks like, the tools it has and when to use each, when to stop, what to do when unsure (ask or assume), and 2 things it must never do. Write it as clear prose instructions, not a list of rules in capitals.
Why it works: Claude follows explained instructions better than shouted rules, so prose with reasons beats MUST and NEVER.
Design the tools
Design the tools for an agent that [does the job]. For each tool: a name, a description that says exactly when to use it and when not to, the input schema with types, and an example call. Use the fewest tools that cover the job.
Why it works: The tool description is the model's only guide to when to call it. Vague descriptions cause most wrong tool calls.
Debug a failed run
Here's a transcript where my agent failed at [goal]. Find the first step where it went wrong, explain why the model made that choice (unclear instruction, missing tool, misleading tool output?), and propose the smallest prompt or tool change that would have prevented it. [paste the transcript]
Why it works: The first wrong step is the root cause; everything after it is fallout.
Write an eval set
Write 15 test cases for an agent that [does the job]: 8 typical requests, 4 edge cases, and 3 it should refuse or hand to a human. For each: the input, what a correct result looks like, and how to grade it automatically.
Why it works: Without a fixed test set, every prompt change is a guess.
Grading rubric for an LLM judge
Write a rubric for grading [type of output] from my agent: 4 criteria, each scored 1 to 3, with a concrete example of a 1 and a 3. Then write the exact prompt a model should use to grade one output against this rubric and return JSON.
Why it works: Examples of a 1 and a 3 keep a model grader consistent from run to run.
Fix a prompt from bad outputs
Here's my prompt and 3 outputs I didn't like. For each, explain what in the prompt caused it. Then rewrite the prompt by fixing the underlying instruction, not by adding a rule per failure. [paste the prompt] [paste the outputs]
Why it works: A rule per failure piles up until the prompt contradicts itself.
Plan a multi-step agent
Break [goal] into the steps an agent would take. For each: the input it needs, the tool or model call, how it knows the step worked, and what to do if it fails. Mark which steps can run in parallel.
Why it works: Success checks per step are what let an agent recover instead of carrying on wrong.
Guardrails review
Review this agent design for ways it could go wrong: actions that are hard to undo, data it could leak, instructions inside tool results it might follow, and loops it could get stuck in. Suggest a guardrail for each. [describe or paste the design]
Why it works: Instructions hidden in web pages or files are the risk agents most often miss.
Turn a manual task into an agent spec
I do this task by hand: [describe the steps]. Turn it into an agent spec: the goal, inputs, tools needed, the steps, where a human should approve, and how we'd know it worked.
Why it works: Writing down where a human approves keeps the agent from acting where it shouldn't.
Summarize a long agent trace
Summarize this agent run for a teammate: the goal, what it did in order (one line per step), what worked, where it struggled, and the final state. Under 200 words. [paste the trace]
Why it works: Long traces burn context. A 200-word summary is enough to pick the work up in a fresh chat.
The Limit Saver prompts: make Claude's usage limit last longer
Every ClaudeKit pack includes these five. Claude re-reads the whole conversation for every reply, so long chats use up your limit fastest. These keep chats short and let you restart without losing your place.
Handoff brief
Before I switch chats, write a handoff brief that someone with no context could pick up cold. Include: 1. The goal, in one or two sentences 2. Decisions made so far, and why 3. What's done 4. What's left, in order 5. Key facts, constraints, names, numbers, and any code we settled on 6. Anything I said I don't want Be complete but brief, and use headings.
Why it works: Run it at about 90%. At 100% you can't ask for anything, and a new chat with this brief costs far less than the long one.
Compress and restart
This chat is getting long. Write a restart prompt I can paste into a new chat: the goal, decisions so far, the current state, open questions, and the exact next step. Under 300 words. Leave out anything we've settled and won't revisit.
Why it works: Claude re-reads the whole chat for every reply, so message 50 costs about 10x message 5. Restarting short resets that cost.
Plan before building
Before writing anything, give me a numbered plan (at most 7 steps) and the questions you need answered. Wait for my OK before you start.
Why it works: A wrong first draft is the most expensive message in a chat. Agreeing on the plan first is cheap.
Short answers mode
For the rest of this chat: answer in under 120 words unless I ask for more. No preamble, no recap, no closing summary.
Why it works: Long answers become long history that every later message pays for.
Only what changed
When you change code or text from earlier in this chat, show only the changed parts, with enough surrounding lines to place them. Don't repeat unchanged sections.
Why it works: Re-printing a whole file to change three lines doubles the chat's length for nothing.
FAQ
How do I write a good system prompt for a Claude agent?
Cover the agent's role in one sentence, what good output looks like, each tool and when to use it, when to stop, what to do when unsure, and the few things it must never do. Write it as prose with reasons. The "Write a system prompt" prompt on this page drafts one for you.
What makes a good tool description for an agent?
It says exactly when to use the tool and when not to, with typed inputs and an example call. The model has nothing else to go on, so vague descriptions cause most wrong tool calls.
How do I evaluate an AI agent?
Write a fixed set of test cases (typical requests, edge cases, and ones it should refuse or hand to a human), each with what a correct result looks like and how to grade it. Then grade with a rubric that has examples of good and bad scores.
Are these prompts free?
Yes. All 15 are free to copy here, and free inside ClaudeKit, a Chrome extension that keeps them one tap away on claude.ai.
Free Chrome extension
Keep this pack one tap away inside Claude.
ClaudeKit adds a prompt library to claude.ai, with this pack ready to insert, and shows your session and weekly usage with the exact reset time. Free, no account.
Add to Chrome, free