AGENTS.md

Guidelines for AI agents working in this repository.

Repository Overview

This repository contains Agent Skills for AI agents following the Agent Skills specification. Skills install to .agents/skills/ (the cross-agent standard).

• Name: PMM-Skills
• GitHub: mia-ppp/PMM-Skills
• License: MIT

Repository Structure

PMM-Skills/
├── skills/                # Agent Skills
│   ├── _shared/           # Shared references skills load (evidence gaps, segment selection, messaging examples)
│   └── skill-name/
│       └── SKILL.md       # Required skill file
├── tools/
│   ├── clis/              # Zero-dependency Node.js CLI tools (61 tools)
│   ├── composio/          # Composio integration layer (quick start + toolkit mapping)
│   ├── integrations/      # API integration guides per tool
│   └── REGISTRY.md        # Tool index with capabilities
├── CONTRIBUTING.md
├── LICENSE
└── README.md

Build / Lint / Test Commands

Skills are content-only (no build step). Verify manually:

• YAML frontmatter is valid
• name field matches directory name exactly
• name is 1-64 chars, lowercase alphanumeric and hyphens only
• description is 1-1024 characters

CLI tools (tools/clis/*.js) are zero-dependency Node.js scripts (Node 18+). Verify with:

node --check tools/clis/<name>.js   # Syntax check
node tools/clis/<name>.js           # Show usage (no args = help)
node tools/clis/<name>.js <cmd> --dry-run  # Preview request without sending

Agent Skills Specification

Skills follow the Agent Skills spec.

Required Frontmatter

---
name: skill-name
description: What this skill does and when to use it. Include trigger phrases.
---

Frontmatter Field Constraints

Name Field Rules

• Lowercase letters, numbers, and hyphens only
• Cannot start or end with hyphen
• No consecutive hyphens (--)
• Must match parent directory name exactly

Valid: acquisition-conversion, email-sequence, marketing-experimentation Invalid: Page-CRO, -page, page--cro

Optional Skill Directories

skills/skill-name/
├── SKILL.md        # Required - main instructions (<500 lines)
├── references/     # Optional - detailed docs loaded on demand
├── scripts/        # Optional - executable code
└── assets/         # Optional - templates, data files

Writing Style Guidelines

Structure

• Keep SKILL.md under 500 lines (move details to references/)
• Use H2 (##) for main sections, H3 (###) for subsections
• Use bullet points and numbered lists liberally
• Short paragraphs (2-4 sentences max)

Tone

• Direct and instructional
• Second person ("You are a conversion rate optimization expert")
• Professional but approachable

Formatting

• Bold (**text**) for key terms
• Code blocks for examples and templates
• Tables for reference data
• No excessive emojis

Clarity Principles

• Clarity over cleverness
• Specific over vague
• Active voice over passive
• One idea per section

House Style

These rules apply to skill files and to every deliverable a skill produces.

• No em dashes. Use a period, comma, colon, or parentheses instead.
• Sentences span two lines at most. Split any longer sentence into two.
• Bullets lead with the outcome. State the result first, then how to get it.

Description Field Best Practices

The description is critical for skill discovery. Include:

1. What the skill does
2. When to use it (trigger phrases)
3. Related skills for scope boundaries

description: When the user wants to optimize conversions on any marketing page. Use when the user says "CRO," "conversion rate optimization," "this page isn't converting." For signup flows, see acquisition-conversion.

Evidence gaps

Evidence decides how an item is labeled, never whether it is included. The full rule lives in one place: skills/_shared/evidence-gaps.md. It covers the Sourced / Assumed / Gap labels, the "Inputs and assumptions" note every deliverable opens with, relevance cuts, customer-facing copy, and the ranked Evidence gaps lists.

• Edit the rule only in that file. It ships with installed skills and the eval harness loads it; this file does neither.
• Load it from every skill that produces a deliverable. Reference it as ../_shared/evidence-gaps.md.
• Load ../_shared/segment-selection.md too from any skill that chooses an audience, ICP, beachhead, or wedge.

Production runs

A production run is a skill producing a real deliverable, such as a GTM plan for an assignment.

• Never edit skill files during a production run. Finish the run with the skills as they are.
• Report skill problems separately, after the run. Name the skill, what the run exposed, and the proposed fix.
• Test skill fixes with a smoke test, not in the run. Production runs do not grade themselves with a pass or fail step.

Testing skill changes

• Check every skill change with an in-session smoke test. Run the changed skill once on one of its eval prompts from evals/evals.json, save the output to scratch/, and hand it to the user to review.
• Run the API eval harness only when the user asks. It is optional. No commit, merge, push, or rollout waits on it.

Git Workflow

Branch Naming

• New skills: feature/skill-name
• Improvements: fix/skill-name-description
• Documentation: docs/description

Commit Messages

Follow the Conventional Commits specification:

• feat: add skill-name skill
• fix: improve clarity in acquisition-conversion
• docs: update README

Pull Request Checklist

• name matches directory name exactly
• name follows naming rules (lowercase, hyphens, no --)
• description is 1-1024 chars with trigger phrases
• SKILL.md is under 500 lines
• No instruction drops content for lacking proof (see Evidence gaps)
• Output format ends with an Evidence gaps section
• No sensitive data or credentials

Tool Integrations

This repository includes a tools registry for agent-compatible marketing tools.

• Tool discovery: Read tools/REGISTRY.md to see available tools and their capabilities
• Integration details: See tools/integrations/{tool}.md for API endpoints, auth, and common operations
• MCP-enabled tools: ga4, stripe, mailchimp, google-ads, resend, zapier, zoominfo, clay, supermetrics, coupler, outreach, crossbeam, introw, composio
• Composio (integration layer): Adds MCP access to OAuth-heavy tools without native MCP servers (HubSpot, Salesforce, Meta Ads, LinkedIn Ads, Google Sheets, Slack, etc.). See tools/integrations/composio.md

Registry Structure

tools/
├── REGISTRY.md              # Index of all tools with capabilities
└── integrations/            # Detailed integration guides
    ├── ga4.md
    ├── stripe.md
    ├── rewardful.md
    └── ...

When to Use Tools

Skills reference relevant tools for implementation. For example:

• referral-program skill → rewardful, tolt, dub-co, mention-me guides
• analytics-tracking skill → ga4, mixpanel, segment guides
• email-sequence skill → customer-io, mailchimp, resend guides
• paid-ads skill → google-ads, meta-ads, linkedin-ads guides

For tools without native MCP servers (HubSpot, Salesforce, Meta Ads, LinkedIn Ads, Google Sheets, Slack, Notion), Composio provides MCP access via a single server. See tools/integrations/composio.md for setup and tools/composio/marketing-tools.md for the full toolkit mapping.

Skill Categories

See README.md for the current list of skills organized by category. When adding new skills, follow the naming patterns of existing skills in that category.

Claude Code-Specific Enhancements

These patterns are Claude Code only and must not be added to SKILL.md files directly, as skills are designed to be cross-agent compatible (Codex, Cursor, Windsurf, etc.). Apply them locally in your own project's .claude/skills/ overrides instead.

Dynamic content injection with !command``

Claude Code supports embedding shell commands in SKILL.md using !`command` syntax. When the skill is invoked, Claude Code runs the command and injects the output inline: the model sees the result, not the instruction.

Most useful application: auto-inject the product marketing context file

Instead of every skill telling the agent "go check if .agents/product-marketing-context.md exists and read it," you can inject it automatically:

Product context: !`cat .agents/product-marketing-context.md 2>/dev/null || echo "No product context file found. Ask the user about their product before proceeding."`

Place this at the top of a skill's body (after frontmatter) to make context available immediately without any file-reading step.

Other useful injections:

# Inject today's date for recency-sensitive skills
Today's date: !`date +%Y-%m-%d`

# Inject current git branch (useful for workflow skills)
Current branch: !`git branch --show-current 2>/dev/null`

# Inject recent commits for context
Recent commits: !`git log --oneline -5 2>/dev/null`

Why this is Claude Code-only: Other agents that load skills will see the literal !`command` string rather than executing it, which would appear as garbled instructions. Keep cross-agent skill files free of this syntax.