Where do I write an agent's prompt in Agent Studio?#
In the LLM (Claude) step. Click it on the canvas and you get three fields that matter: Prompt, System prompt and Model. Use {{in}} inside the prompt to drop in whatever the step before it produced.
| Field | What goes in it | Example from the templates |
|---|---|---|
| System prompt | Who the model is and the rules it never breaks | You are a B2B sales qualifier. Be direct and concrete. |
| Prompt | The task, the output shape and {{in}} | Score this lead 1 to 10 on fit and intent. Output JSON. |
| Model | The cost and speed tier for this step | Leave on the platform default until you need more |
| Draft safeguards | Optional checks for SEO titles and contact copy against your approved business facts | SEO titles and descriptions |
What makes a good AI agent prompt?#
A good agent prompt reads like a brief to a sharp new hire: who they are, what to do, what to hand back, and what to do when they are stuck.
Anthropic gives the same advice for Claude, the model behind every LLM step:
“Think of Claude as a brilliant but new employee who lacks context on your norms and workflows. The more precisely you explain what you want, the better the result.”
- Give a role. "You are a customer-experience manager" sets vocabulary and judgment in one line.
- Name the task in one sentence. Put the action verb first: score, draft, summarize, extract.
- Show the exact output. Paste the JSON shape with every key, or set a word limit for prose.
- Supply the facts. Paste your prices, service area and policies, or wire in Knowledge Search.
- Say what to do when the answer is missing. The FAQ Concierge template tells the model to say so plainly and never invent information.
- Ban the phrases you hate. The Review Responder system prompt forbids "I apologize for the inconvenience".
SYSTEM PROMPT
You are the intake coordinator for {business}. Be direct and specific.
Use only the facts provided. If a fact is missing, write "unknown".
PROMPT
Read the message below and return STRICT JSON only:
{ "intent": "quote | booking | support | other",
"urgency": "today | this_week | later",
"summary": "one sentence",
"reply": "under 60 words, in our voice" }
Message:
{{in}}How do I write prompts for a multi-step agent?#
Give each step one job and make every step carry the earlier findings forward. The AI Visibility Growth Team template chains five Claude steps: a brand strategist, a demand analyst, a competitive analyst, a growth lead and a content lead. Each one adds its own JSON key and returns everything it received unchanged, so the last step assembles the full plan.
How do I test a prompt before going live?#
Wire it with a free dry run, then judge the words on real examples. Run test checks every step and branch without spending credit; model steps return simulated results there, so switch on Run for real when you want to read Claude's actual output.
Test a prompt in five minutes
Step 1: Open a template close to your job
Start from a working prompt instead of a blank one. Lead Qualifier, Review Responder and FAQ Concierge are strong starting points.Browse templatesStep 2: Collect five real inputs
Pick an easy case, a hard case, a messy one, one missing key facts, and one the agent should refuse.Step 3: Run each one for real
Turn on Run for real and read every output against what your best employee would have written.Step 4: Fix the prompt, then save a version
Tighten the rule that failed, rerun the five, and press Save version once all five pass.Step 5: Review live runs weekly
Open Runs, read a handful of results, and add any new failure to your test set.Open Runs