Agent Personas

Connect your agents to the skills their job needs.

A catalogue of 48 personas, 49 specialised skill bundles and 308 reusable skills for building dedicated agents, sub-agents and bots. Choose a role, load its methods, and use the agent platform you already have.

Choose a persona · Try it · Browse all roles and bundles · Platform compatibility · Evidence

The connection between agents and skills

An agent gives you a model, tools and somewhere to run work. A skill gives it a method for a particular task. Putting a folder of skills beside an agent still leaves you to decide what its job is, which methods belong together, when to use them and what a finished result should contain.

Agent Personas packages that connection. Each persona defines a role and working expectations, then connects it to a specialised bundle of reusable skills. You can give an agent a research job with source-verification methods, or an implementation job with testing and review methods, without assembling the instructions from scratch each time.

Your agent platform
  │  provides the model, tools, memory and permissions
  ▼
Persona
  │  defines the job, decisions, boundaries and completion criteria
  ▼
Specialised skill bundle
  │  selects the complementary methods the role needs
  ▼
Focused skills
     guide individual tasks and checks

The persona provides direction; its skills provide procedures. Your platform remains in charge of execution and permissions.

What that looks like in practice

Take the Software Implementation Specialist. Its job is to deliver bounded features and fixes with verification. Its bundle connects that job to:

A refactoring persona has a different emphasis: establish compatibility tests before restructuring, preserve observable behaviour and measure simplification without deleting functionality to hit a target. A research persona instead coordinates source checking, synthesis and citations.

These are written working methods. They give you a starting point to inspect and adapt, rather than a promise that a model will follow every instruction.

Choose your first persona

Start with the outcome you need. The coding roles sit alongside research, design, product and everyday-use roles in the same catalogue.

Browse all personas, bundles and skills →

Browse all personas and skill bundles

The directory below covers all 48 persona role contracts and 49 skill bundles. A persona defines the job, boundaries and completion criteria; a bundle selects methods without assigning a role. Each row links to both separately. Counts are declared member skills, not unique skills across the catalogue; shared skills are counted in each package that includes them.

Paired personas and bundles currently have identical membership. Expand a group to inspect the linked skills for each entry. The standalone knowledge-librarian-bundle has no matching persona. Conditional members are included in counts but activate only when relevant and authorised; they grant no dispatch permission.

Software, architecture and data

Design, research and media

Product, delivery and operations

Agents, models and knowledge

Learning and everyday planning

Try a persona in a few minutes

You need Git and Python 3.10+ for the optional staging helper. Reading and using the instruction files does not require Python.

1. Get the catalogue and verify it

git clone https://github.com/Sahil-SS9/agent-personas.git
cd agent-personas
python3 scripts/verify.py

The verifier checks package structure, file integrity, membership and local references. It does not contact a model or require credentials.

2. Stage one role in a new workspace

python3 scripts/try_profile.py \
  --select persona:software-implementation-specialist \
  --harness codex \
  --dest ../implementation-agent-trial

Replace codex with claude, hermes, openclaw, commandcode, pi or opencode as appropriate. The helper copies the persona and its member skills into a new directory. It refuses to overwrite an existing destination and does not change your live profiles.

3. Give your agent a concrete job

Open your agent in the staged workspace, explicitly load PERSONA.md and its relevant member skills, then give it a bounded task. For example, in a disposable copy of your project:

Read PERSONA.md and load the relevant member skills. Investigate this failing test, implement a scoped fix and report the checks you ran. Do not merge or deploy.

Staging files does not itself load them into a model. Use your platform’s permission controls for workspace safety. The getting-started guide covers activation and optional CLI smoke runs in more detail.

Use the pieces your setup needs

You can adapt a persona for a main agent, a scoped sub-agent or a bot. Configure the account, model, tools, memory and channels through your chosen platform. This repository does not provision those resources or grant permission to publish, deploy or access production systems.

Keep shared methods in skills rather than copying them into every persona. That lets you reuse a method across roles and review changes without maintaining several competing copies. Install one version per intended scope to avoid discovery collisions.

Bring your own agent platform

The instruction cores are portable Markdown. Installation guidance is separate, so the roles do not require one vendor’s runtime or delegation API.

The compatibility matrix distinguishes documented installation routes, filesystem staging, native discovery and agent execution. A passing staging test is not a claim that every role has been run successfully in every client. The JSON persona and bundle manifests describe this catalogue’s composition; they are not universal native-agent schemas.

Methods you can inspect, evidence you can check

The catalogue includes source attribution, method references and per-skill evidence cards. Coding methods draw on selected material about legacy-code seams, preparatory refactoring, API compatibility, testing and review. The coding evidence notes explain what the comparisons covered and where the limits remain.

We test package integrity separately from agent behaviour. Some behavioural comparisons showed useful changes in the order and depth of checks; others had equal outcomes or exposed regressions. Release status does not turn those observations into universal accuracy or speed claims. The evidence guide keeps those distinctions visible.

The private builder, raw research and hidden evaluation suite are not distributed. The consumer verification and staging scripts run independently of that tooling.

Release status

Release 0.5.0 adds the Reverse Engineering Specialist and expands the Frontend UX Engineer. The newest instruction revisions were accepted by the owner without a further benchmark; evidence cards apply to the versions they record. Package integrity checks are separate from behavioural claims. See the release history and evidence guide.

Adapt it, then share what you find

Pick a role you have work for, try it on a bounded task and adjust it to your workflow. Reports showing a missed check, an unclear trigger or a redundant method are especially useful. Include the persona, platform, task and observed behaviour so others can reproduce the issue.

Contributing · Security · Release history

Licence and acknowledgements

Original content and consumer code are MIT licensed. Attribution and applicable third-party terms are recorded in THIRD-PARTY-NOTICES.md and package notices.