Contributing an Agent Template

This guide explains how to contribute a new agent template to the pm-claude-skills library.

What is an agent template?

An agent template is a runnable workflow that combines existing skills, connectors, and subagents into a single end-to-end task. Following the architecture Anthropic introduced for financial services agent templates on May 5, 2026.

Examples of agent templates that would belong in this repo:

PM Discovery Agent — combines discovery-interview-guide + user-interview-synthesis + assumption-mapper with Granola/Notion connectors
Legal Contract Review Agent — combines contract-review + nda-analyser + compliance-checklist with Google Drive connector
Sales Pursuit Agent — combines sales-battlecard + discovery-call-prep + proposal-writer + account-plan with Salesforce/Gong connectors

Required structure

Every agent template needs these files:

templates/your-agent-name/
├── README.md                  # What it does, install, usage
├── AGENT.md                   # Agent definition (system prompt + tool list)
├── orchestrate.sh             # Orchestration script
├── skills/                    # Skills used (linked from main library)
│   ├── README.md
│   └── [skill-name]/SKILL.md
├── subagents/                 # Specialised subagents
│   └── [subagent-name].md
├── connectors/                # Data source configurations
│   ├── README.md
│   └── [system].example.json
├── examples/                  # Input and output examples
│   ├── input-example.md
│   └── output-example.md
└── tests/
    └── smoke-test.md

Naming conventions

Folder name: Use kebab-case, descriptive of the workflow (e.g., pm-sprint-agent, legal-contract-review-agent, sales-pursuit-agent)
AGENT.md: Always exactly this name (with caps) so it's easily findable
Subagent files: kebab-case in subagents/, ending in .md (e.g., capacity-analyst.md)
Connector files: lowercase, with .example.json for the template version (e.g., linear.example.json)

Quality bar for new templates

Before submitting a PR, verify:

README.md explains what the agent does in the first paragraph (no more burying the lede)
AGENT.md has a complete system prompt with explicit step-by-step instructions
At least 2 skills from the main library are referenced (otherwise it's just a skill, not a template)
At least 1 subagent is defined for analysis the skills can't do alone
At least 1 connector with a working example config
orchestrate.sh runs without errors in --dry-run mode
Smoke test passes (documented in tests/smoke-test.md)
Example input AND example output are provided
Honest limitations section in the README — what the agent doesn't do well
No credentials in any committed file — credentials must come from environment variables

What makes a good agent template (vs a bad one)

Good agent templates:

Solve a specific, recurring professional workflow end-to-end
Have clear separation between skills (output formats), connectors (data access), and subagents (specialised analysis)
Work without modification for a typical team in the target profession
Include honest limitations and caveats

Templates that get rejected:

Wrap a single skill with no real orchestration ("just call the skill")
Combine unrelated skills with no coherent workflow
Hardcode credentials or organisation-specific data
Don't include working examples
Don't include subagents (just skills + connectors isn't a template)

How to submit a PR

Fork the pm-claude-skills repo
Create your template in templates/your-agent-name/
Run the smoke test successfully
Commit your changes with a clear message: feat: add [agent name] template
Open a PR with this description:
- What this template does (1 paragraph)
- Which skills it uses (list)
- Which connectors it requires (list)
- Which subagents it defines (list with one-line descriptions)
- Smoke test result (paste the output)

PRs get reviewed within 5-7 days. The review focuses on the quality bar above, not personal style — clean templates that meet the bar get merged.

What you get for contributing

Credit in the main README under the contributing section
Mention in the next Medium article in the Claude Skills series
Maintainer access to your template — you can update it directly without needing review for minor changes after the first merge

Questions?

Open a discussion before you start building if your template doesn't fit cleanly into the structure above. It's much easier to align early than to rework after.

4.9 KiB Raw Blame History