Michal Sitarzewski's open-source repository msitarzewski/agency-agents contains more than 140 agent personas across 12 divisions.

Each persona is a Markdown prompt that defines a role, its scope, expected outputs, and common failure modes.

The useful part of the repository is its prompt structure: the personas specify decision frameworks, operating boundaries, and deliverables instead of stopping at a role label.

Beyond an "act as" prompt

Telling an LLM "Act as a CFO" or "You are a senior lawyer" supplies a role but leaves the task, constraints, and output format unspecified.

Looking across the files, I found four recurring kinds of instruction. These are my labels for the pattern, not the repository's formal taxonomy:

Role and tone

The role's perspective and tone, such as cautious legal review or aggressive growth experimentation.

Responsibilities and limits

What the role should handle and what falls outside it.

Expected output

The requested format, such as a unit-economics model, threat matrix, or product requirements document.

Common mistakes

Failure modes the prompt tells the model to check before returning an answer.

Using several personas on one question

One way to use the catalog is to send the same proposal to several personas and compare their responses:

  1. Run separate sessions: Open three agent sessions in Claude Code, Cursor, or Antigravity, using the Chief Strategy Officer, Head of Growth, and Principal Systems Architect personas.
  2. Use the same input: Give each session the same product pivot or architecture proposal.
  3. Compare the outputs: The prompts direct the architect toward technical debt, the growth lead toward acquisition, and the strategy role toward defensibility. Their answers still require fact-checking and human judgment.

This is a structured way to collect several model-generated perspectives. It is not a substitute for experts, and the personas do not add knowledge beyond the underlying model and supplied context.

Loading the prompts into other tools

Because the library consists of Markdown files, its prompts can be adapted to other developer environments. The paths below are examples of possible adaptations, not canonical installation instructions from the repository; each tool's current configuration and precedence rules still need to be checked.

  • Claude Code or a terminal harness: Adapt a persona as an agent definition or explicit system prompt where the tool supports one.
  • Cursor or Windsurf: Rewrite the relevant instructions as workspace rules in the format supported by that version of the editor.
  • Aider: Extract project conventions into a file the session is explicitly told to read.
  • Custom harnesses: Register selected instructions as a local role or skill after removing assumptions that do not fit the task.

What the open repository provides

The repository can be inspected and modified without depending on a hosted prompt library:

  • Inspectable: Every instruction is visible in the repository.
  • Forkable: The files can be adapted to a company's terminology, stack, and compliance requirements.
  • Compatible with local use: When paired with a fully local model and toolchain, prompts and execution logs can remain on the local machine.

Primary References & Links