A persistent AI operating system built entirely from markdown files.
Seven architectural layers. Thirteen specialized agents. A 5-deliberator quality gate (agent-council) for tier-1 artifacts. A recursive improvement loop that compounds every session. Zero dependencies.
| Agents | Skills | Guardrails | Workflows | Learnings | Dependencies |
|---|---|---|---|---|---|
| 13 | 12 | 8 | 4 | 100+ | 0 |
AI has three ceilings that no amount of model intelligence fixes.
| Ceiling | Without Agent Prime | With Agent Prime |
|---|---|---|
| Context | Every session starts from zero | 42+ decisions, 14+ voice rules, 11+ design patterns -- all persistent |
| Consistency | Same mistakes, different sessions | Every correction becomes a permanent rule across all agents |
| Throughput | One task, one thread | Parallel workstreams: research, analysis, writing, building |
Agent Prime is seven layers of persistent infrastructure -- markdown files that encode your thinking model, your memory, your standards, and a recursive loop that gets smarter every session. Every correction you make becomes a permanent rule, applied automatically across all agents, in every future session.
Real pipeline outputs. One sentence in, structured analysis out.
Pipeline: Scout --> Industry Analyst --> Investment Analyst Output: 1,328 lines | 13 frameworks | 91 citations
8-layer value chain decomposition, bottleneck analysis with severity scores, 6 deployment scenarios with probability weights, 4-lens investment valuation, stress-tested watchlist with exit rules.
Raw AI on the same question: ~800 words of generic overview.
Pipeline: Problem Framing --> Discovery Research --> Competitive Analysis --> Metric Design --> Spec Writing --> Narrative Building Output: 2,399 lines across 6 artifacts | 30+ frameworks | 70+ citations
Six PM skills chained together. Each artifact feeds the next -- problem framing shapes research, research informs competitive analysis, analysis drives metrics, metrics anchor the spec, spec feeds positioning. One sentence of input produced a complete product strategy across 6 interconnected documents.
Pipeline: Scout --> Synthesizer --> Planner --> Builder Output: 1,342 lines across 4 artifacts | 10 sourced signals | 4-phase implementation
Scout gathered 10 signals (MIT, Stanford, HBS studies + Stripe, Figma, Linear examples). Synthesizer produced a thesis: "The PM Skill Stack is Inverting." Planner designed a 4-layer team OS. Builder produced a phased implementation plan with day-by-day scheduling.
| Raw AI | Agent Prime | |
|---|---|---|
| Input | Same question | Same question |
| Output | 500-800 words | 1,300-2,400 lines (30-50 pages) |
| Frameworks applied | 0 | 13-30+ |
| Evidence citations | 0 | 70-91 |
| Actionable deliverables | "Do more research" | Watchlists, specs, implementation plans |
Each layer is independently useful. Together, they compound.
| Layer | What It Does | Key Files |
|---|---|---|
| 01 Identity | Encodes who you are -- thinking patterns, judgment heuristics, epistemic guardrails | shared/context.md |
| 02 Memory | Every correction compounds. Fix something once, it propagates everywhere. | shared/learnings.md, shared/dependency_map.md |
| 03 Agents | 13 specialized agents -- scout, synthesizer, writer, planner, builder, and more | agents/*/prompt.md |
| 04 Orchestration | Registry tracks work, dispatch sequences tasks, briefings surface priorities | shared/registry.json, prime/dispatch.md |
| 05 Skills | 12 PM methodologies -- competitive analysis, metric design, pricing, GTM, and more | shared/toolkits/skills/ |
| 06 Craft | Design system + templates that produce publication-quality HTML from markdown | agents/builder/templates/ |
| 07 The Loop | The recursive mechanism -- learnings propagate, agents chain, the system audits itself | Emerges from layers 1-6 |
See ARCHITECTURE.md for the full system design.
git clone https://github.com/avyayalaya/agent-prime.git
cd agent-prime
# Point your AI assistant at it:
# Claude Code — auto-loads CLAUDE.md
# Copilot — auto-loads .github/copilot-instructions.md
# Any other LLM — load CLAUDE.md as your system promptWant the guided setup? Run python meta/scripts/first_run.py for interactive onboarding (preview, quick trial, or full setup).
Verify everything works:
python meta/scripts/verify_setup.pySee QUICKSTART.md for the detailed walkthrough.
| Test | Score | What It Measures |
|---|---|---|
| PM Skills (competitive analysis) | 93.3/105 (88.9%) | Output quality vs. skill-agnostic rubric |
| Baseline (no system) | 47/105 (44.8%) | Same task, vanilla Claude |
| Anthropic PM Skill | 81/105 (77.1%) | Same task, Anthropic's built-in PM skill |
Agent Prime's skill-powered output scores 2x baseline and 21% above Anthropic's own PM skill on the same structured rubric.
More benchmarks in development: session continuity, correction compounding, throughput multiplier, cross-LLM portability.
agent-prime/
├── CLAUDE.md ← System instructions (auto-loaded by Claude Code)
├── .github/copilot-instructions.md ← Copilot instructions (auto-injected by VS Code)
├── agents/ ← 13 agent prompts (prime, scout, writer, planner, ...)
├── shared/
│ ├── context.md ← Your identity, goals, and voice rules
│ ├── registry.json ← All work items (single source of truth)
│ ├── learnings.md ← Accumulated corrections (append-only)
│ ├── dependency_map.md ← Change propagation registry
│ └── toolkits/skills/ ← 12 PM methodologies (~1,200 lines each)
├── prime/
│ ├── dashboard.md ← Auto-generated status view
│ ├── briefing.md ← Session-start priorities
│ ├── dispatch.md ← Task queue with successor rules
│ └── config.json ← Agent cadences and goal mappings
├── examples/
│ ├── ai-robotics-industry/ ← Pipeline: investment analysis (1,328 lines)
│ ├── product-strategy-figma/ ← Pipeline: 6-skill product strategy (2,399 lines)
│ └── pm-team-performance/ ← Pipeline: team transformation (1,342 lines)
├── meta/scripts/ ← Python automation (briefing, dashboard, integrity)
└── getting-started/ ← Guided projects (20-25 min each)
Agent Prime is plain markdown. No vendor lock-in.
| Environment | How It Loads | Setup |
|---|---|---|
| Claude Code | Auto-reads CLAUDE.md at session start |
Clone and go |
| GitHub Copilot | Auto-injects .github/copilot-instructions.md |
Clone and go |
| Gemini CLI | Load CLAUDE.md as system context |
Manual load |
| Any LLM | Feed CLAUDE.md + relevant agent prompts |
Manual load |
The system prompt file is the same regardless of model. Agents, skills, and memory are model-agnostic.
See CONTRIBUTING.md. New skills, agent improvements, and documentation fixes welcome. Keep PRs focused.
MIT. See LICENSE.