A fully opinionated set of documentation to shape AI agent builds and keep complexity contained.
PocketSWE tells AI agents how to build software. Drop it into a greenfield project and agents will follow consistent architectural patterns, verification procedures, and work protocols.
It is packaged the way coding agents load context:
AGENTS.md— a lean, always-on contract holding the non-negotiable architectural invariants: the spine, the dependency direction, and the prohibited-complexity list. Auto-discovered by the many tools that support the AGENTS.md standard (Codex, Cursor, Copilot CLI, Gemini CLI, Claude Code, and more).- Skills (
.agents/skills/) — phase-specific entry points loaded on demand via progressive disclosure: only each skill's name and description sit in context until a matching task loads the body, which directs the agent to the full source document indocs/. docs/— the full-text sources and single source of truth: each skill references them directly, andAGENTS.mdlinks them as the portable fallback for tools without skill support.
These documents encode hard-won opinions about:
- Architecture - Onion architecture with explicit dependency inversion, SOLID principles, and high-cohesion modules
- Verification - End-to-end feature verification against a defined "North Star"
- Work Protocol - A structured workflow from discovery through integration
AI agents are capable builders, but they benefit from constraints. Without guidance, agents make inconsistent architectural decisions, skip verification, or produce code that becomes difficult to maintain as it grows.
PocketSWE locks in the boring decisions upfront so agents can focus on solving your actual problem. And because always-on rules are expensive — every line of AGENTS.md rides along on every task — only the invariants stay resident; each procedure loads only when its phase of work begins.
| Path | Loaded | Purpose |
|---|---|---|
AGENTS.md |
Always | Spine, dependency direction, prohibited complexity, pointers to skills and docs |
.agents/skills/planning-work/ |
When planning | Entry point for the work protocol — references docs/work-protocol.md |
.agents/skills/building-features/ |
When implementing | Entry point for feature construction — references docs/architecture.md |
.agents/skills/verifying-features/ |
When verifying | Entry point for North Star verification — references docs/verification.md |
docs/architecture.md |
On demand | Full architecture contract (layer rules in detail) |
docs/work-protocol.md |
On demand | Full work protocol |
docs/verification.md |
On demand | Full verification contract |
- Copy
AGENTS.md,.agents/, anddocs/into your project root - Build — tools that support
AGENTS.mddiscover the contract automatically
Agents plan with planning-work, implement with building-features, and accept nothing as done until verifying-features passes. Tools without skill support get the same content through the docs/ links in AGENTS.md.
Claude Code discovers Skills under .claude/skills/ and reads CLAUDE.md. Symlink so the canonical .agents/ files are picked up:
mkdir -p .claude && ln -s ../.agents/skills .claude/skills # skills
ln -s AGENTS.md CLAUDE.md # optional: instruction file- Greenfield projects where you want to establish patterns early
- Teams using AI agents for feature development
- Projects where maintainability matters more than speed-to-first-commit
- Brownfield projects with established (different) patterns
- Quick prototypes where structure is overhead
- Projects where the existing architecture should be preserved
MIT License. See LICENSE.
Issues and pull requests welcome at github.com/TypeScale-co/PocketSWE.