How to Write a PydanticAI Agent Spec from a Role List
PydanticAI specs fail when they paste sk- and invent gpt-4o. The role list named intake and parse_ticket. Stop there.
The matching generator is the PydanticAI Agent Spec from a Role List (No Invented API Keys or Tools) (Pydanticai Agent Spec From Role List No Invented Api Keys Or Tools) prompt. Browse related cards in the PromptDig library (Browse more prompts). When a filled run survives, share the version you actually use (Share a prompt).
Quote PydanticAI version or write UNKNOWN
Write a PydanticAI agent spec from a role list. No invented API keys, model IDs, or tools. Start by filling Inputs, not by asking the model to remember last week's run. If a field is blank, write NONE or NOT IN INPUTS and leave it blank through Generate. The card is built so the model cannot honestly invent a number, owner, URL, or command that you did not paste.
Paste these fields before you hit run:
Role list I lock (names, jobs, tools I allow): [Roles]
Model IDs I lock (or UNKNOWN): [Models]
PydanticAI version I lock (or UNKNOWN): [Version]
Output types I lock: [Types]
Words I must not use: [Banned]
What I must never invent (API keys, extra tools, model IDs): [Never]
Output format: [Format]
Language for comments: [Lang]
Secrets style I lock (env names only): [Secrets]
That inventory is the honesty ledger. Anything that does not appear there is forbidden in the draft. If you catch yourself adding a nice-to-have after the run, you are no longer using the card. You are ghostwriting. Put the extra fact in Inputs and run again.
Wire only named tools
Generate is numbered on purpose. Do not skip a step because the first paragraph looked done. The early steps exist to stop later prose from smuggling claims.
Walk the Generate list in order:
- Honesty ledger: Roles, Models, Version, Types, Secrets, Lang. Forbidden: invented keys, extra tools, extra models.
- Tool lock table: each tool in Roles quoted. Tools not listed stay NOT IN INPUTS.
- Spec: Agent(name, tools, output_type) only from Roles and Types. Model from Models or MODEL UNKNOWN.
- Secrets: env names from Secrets only. Never paste sk- values.
- Refuse list: CrewAI crew.yaml, LangGraph add_node, Haystack YAML, raw API keys.
- Version: quote Version or write VERSION UNKNOWN.
- Never: do not add web_search if Roles omitted it. Do not invent gpt-4o.
- Compliance pass: quote Banned and Never hits. Cut them. Format as Format.
If a step asks for a version lock, quote the version from Inputs in the output. If a step asks for a refuse list, keep the refuse list in the published artifact, not in a sidebar you delete. Reviewers should see what the model was not allowed to do.
Secrets stay as env names not raw keys
Most failures are the same shape: a missing field gets a confident fill. A conversion rate appears. A Gradle task appears. A flash point appears. A caption appears on a job that asked for slide text only. Your review is to search the draft for numbers, names, and commands, then grep Inputs. No match means cut.
Honor the constraints as hard stops, not vibes:
- PydanticAI spec from Roles only. Not CrewAI and not LangGraph.
- Never invent API keys, extra tools, or model IDs.
- Secrets stay env names.
- No emojis.
When the card says not legal advice, not certification, not an exam dump, or not a caption engine, that sentence belongs at the top of the output. Deleting it to look more finished is how you inherit risk.
Refuse CrewAI LangGraph and Haystack dumps
Finish with the compliance pass the prompt already asks for. Quote the banned-word hits. Cut them. Print character counts when the job has a cap. Print word counts when the job has a budget. List gaps as gaps. Five missing facts are more useful than one smooth paragraph.
Tags on the card (pydanticai agent spec, role list lock, no invented api keys) are a reminder of the job shape, not an invitation to wander into a neighboring cluster. If you need a different surface, open a different PromptDig card rather than stretching this one.
Fill the card, then run
Replace every bracket. Run on ChatGPT, Claude, or Gemini. Read the ledger first, then the artifact. If the model invents a commit, KPI, DOI, PEL, bid, or logo, discard the run. Tighten Inputs. Run again. Share the filled card that survived, not the first draft that sounded done.