MOC: Coding Harnesses, Skills & Frameworks

Map of Content for the layers around a coding model: the harness it runs in, the skills that teach it, and the frameworks that impose a workflow (plan, execute, verify, review) on top.

Harnesses (the agent runtime)

• pi.dev — minimal, terminal-first, deeply extensible.
• harness-engineering-openai — OpenAI's account of shipping a product with Codex and no hand-written code; how to shape the environment.
• symphony-openai — turns a Linear board into isolated autonomous runs.
• moc-agent-workbenches — environments that run and coordinate agents (bb, paperclip, dmux).

Skills (packaged know-how)

• anthropic-skills-guide — Anthropic's guide to building skills.
• skill-vs-agents-vs-claude-md — where each kind of instruction belongs.
• moc-best-ai-skills — shortlist of skills worth adopting.
• company-skills-library · skills-managers · openskills — sharing and distributing skills.

Frameworks (workflow on top)

• pstack — Cursor plugin: 23 skills, 21 principles, 22 playbooks; /poteto-mode routes work, verification and multi-model review built in. Skills are SKILL.md, so portable (unofficial Claude Code port: pstack-claude). Review: pstack-review-flavio-copes.
• gsd-core — spec-driven loop (discuss, plan, execute, verify, ship) using fresh-context subagents.
• moc-loop-engineering — loops that prompt agents on a schedule.
• moc-planning-work — planning before delegating.

Pattern

SKILL.md is becoming the common packaging format across Claude Code, Codex and Cursor, so frameworks increasingly ship as portable skill folders rather than tool-specific plugins. What differs is the harness features they lean on (per-role model assignment, subagents, hooks).

---

Grows as harnesses, skills and frameworks earn a place.