Official

# AGENTS.md and Agent Workflow Directory AGENTS.md is a portable convention for repository instructions used by multiple coding agents. AgentAtlas collects public examples and agent workflows that technical teams can evaluate, reproduce, and adapt. ## Write Useful Repository Instructions - **Commands**: include exact setup, development, test, lint, build, and focused-test commands. - **Architecture**: explain important paths, generated code, ownership, and decisions that are hard to infer. - **Conventions**: document repository-specific patterns instead of repeating generic language guidance. - **Boundaries**: identify secrets, protected files, destructive actions, and required approvals. - **Verification**: define the user journey and checks required before work is complete. - **Scope**: use nested AGENTS.md files when a subdirectory needs more specific guidance. ## What a Complete Listing Includes - **Outcome**: the job the agent completes and the user or team it serves. - **Triggers and inputs**: the event that starts work plus the required files, records, messages, or API payloads. - **Actions and outputs**: the systems the agent reads or changes and the artifact it produces. - **Integrations**: confirmed APIs, MCP servers, SaaS products, databases, and required credentials. - **Human control**: approval gates, escalation paths, destructive actions, and sensitive-data restrictions. - **Operations**: retries, idempotency, status reporting, logging, support ownership, and known failure modes. - **Evidence**: current documentation, a demonstration, a public repository, pricing, or another link readers can verify. ## Review Before Adopting Directory inclusion and paid placement are not security audits or endorsements. Confirm permissions, data retention, licensing, pricing, support, and current product behavior with the maintainer before using an agent with production data. ## Submit a Workflow Read the cross-tool format at [agents.md](https://agents.md/), then publish an entry through the [AgentAtlas submission form](/agents-md/new). Questions, corrections, and partnership inquiries can be sent to [partners@agentatlas.dev](mailto:partners@agentatlas.dev).
# Product Requirements for AI-Assisted Delivery A useful product requirements document aligns people and agents on an outcome, scope, workflows, constraints, and evidence of success. It should reduce ambiguity without turning an uncertain product decision into an assumed technical requirement. ## Minimum Useful Structure - **Outcome**: the user or business change and the metric that will show whether it happened. - **User and problem**: who experiences the problem and what evidence shows it is worth solving. - **Scope**: what is included, excluded, assumed, and deferred. - **Workflows**: starting states, actions, system responses, success paths, and recovery paths. - **Requirements**: functional behavior plus privacy, security, performance, reliability, and accessibility constraints. - **Acceptance criteria**: observable conditions that product, design, engineering, and QA can evaluate. - **Measurement**: primary metric, guardrails, and the instrumentation required. - **Risks and decisions**: open questions with an owner and decision date. ## Prepare the Handoff Link designs, schemas, APIs, analytics events, existing implementation patterns, and decision records. Label confirmed requirements separately from assumptions and proposed solutions. Include commands, environments, test cases, and user-visible checks so an implementation can be verified. [Share a PRD](/prd/new) with the AgentAtlas community.
# CLAUDE.md Project Instructions A project CLAUDE.md gives Claude Code durable, repository-specific context. Use it for commands, architecture, conventions, safety boundaries, and completion checks that apply across sessions. ## What to Include - Install, development, test, lint, build, and migration commands. - Important source directories, ownership boundaries, and sources of truth. - Specific coding, validation, accessibility, security, and review requirements. - Sensitive paths, prohibited actions, and decisions that require human approval. - The checks that must pass before work is considered complete. ## Keep It Effective Prefer concise, concrete instructions that can be verified. Remove information Claude can discover easily, avoid conflicting rules, and link to maintained documentation instead of duplicating long material that will become stale. Project instructions shape model behavior; they do not enforce policy. Use permissions, hooks, protected branches, CI, and application controls for requirements that must always hold. Check the current [Claude Code documentation](https://code.claude.com/docs/en/memory) for supported locations and loading behavior, or [submit a CLAUDE.md example](/claude/new).
# Agent Commerce: UCP and AP2 Agent commerce now has complementary layers. Universal Commerce Protocol (UCP) covers discovery, product and offer data, checkout, and order lifecycle interactions. Agent Payments Protocol (AP2) adds verifiable artifacts for payment intent, purchase terms, and authorization. ## Where Each Protocol Fits - **UCP**: connects agents and businesses across the commerce journey. - **AP2 Intent Mandate**: records the user's request and constraints. - **AP2 Cart Mandate**: captures the concrete purchase terms presented for authorization. - **Verifiable authorization**: signed artifacts help participants reason about consent and preserve an audit trail. - **A2A and MCP**: can provide discovery or service-invocation layers around a commerce workflow. ## Implementation Review Document agent authority, consent, participant identity, credential handling, duplicate-request protection, price and inventory changes, payment failures, refunds, disputes, retention, and audit access. Protocol support does not replace legal, fraud, security, or payment-compliance responsibilities. Use the maintained [UCP guide](https://developers.google.com/merchant/ucp), [UCP specification](https://ucp.dev/), and [AP2 specification, schemas, SDK, and scenarios](https://github.com/google-agentic-commerce/AP2) as primary sources. Pin the versions you test in production.