Skip to main content

Prometheus Skill Pack — Official Product Documentation

Maintained by Travis James, CTO, Prometheus AGS · Licensed MIT

This is the canonical product documentation for the prometheus-skill-pack: a self-improving AI skill execution engine that turns the loop — not the prompt — into your primary unit of work.

Most teams treat an AI coding agent as something you talk to. You type, it responds, you type again. That posture caps out fast. The prometheus-skill-pack is built on a different premise: you design the loop that prompts the agent, the loop remembers what it learned, and the system proposes improvements to its own skills as it discovers better ways to work. The agent executes. You write loops.

That premise is not aspirational. It is implemented across supported AI tools as a four-layer pipeline, three nested loop levels, a Karpathy-pattern knowledge base, a structural anti-sycophancy gate, a Rust toolchain that generates agents, skills, and native tools, and Dynamic Operations that turn eligible generated code into bounded runs with signed evidence.

This documentation explains all of it. Every skill, every tool, every CLI, every script, every hook — documented individually and then collectively, with the architecture diagrams that make the design legible.


How this documentation is organized​

The guide is built in layers. Read it top to bottom the first time; use it as a reference after that.

Foundations — the why and the what​

#PageWhat it covers
01IntroductionWhat the skill pack is, who it is for, the autonomy ladder, and the loop posture
02Metaprompting, PMPO, and KBDThe methodology: metaprompting, Prometheus Meta-Prompting Orchestration, Knowledge-Based Development, and the theory behind them
03Loop ArchitectureThe L0–L3 loop levels, nested loops, loop.json, the loop-tick.sh exit-code contract, feedback sources, escalation, and autonomy gates
04The Four-Layer PipelineZeeSpec → PMPO → OpenSpec → forge-rs, with C4 container diagrams

The substrate — what makes loops compound​

#PageWhat it covers
05The MCP Server SubstrateAll eight MCP servers, the canonical port table, and Firecrawl vs. Tavily
06Memory and Karpathy-Pattern LearningThe three-layer memory architecture, the self-learning engine, and the cross-session write-back sequence
07Sycophancy CorrectionThe eight patterns (S-01–S-08), modes, strictness, MCP tools, the reflection gate, and how this documentation was checked with it

The catalog — every skill​

#PageWhat it covers
08Skills OverviewThe skills model, discovery, the AgentSkills.io standard, and the full category index
09Process & Orchestration SkillsZeeSpec, iterative-evolver, the KBD orchestrator and its child skills, pmpo-elicit, pmpo-outer-loop, pmpo-skill-creator, native-agent, liter-llm-bridge, ideation-mindmap, kbd-evolve
10Learn Domain SkillsThe 12 learn skills (ui-surface, learn-goal, learn-survey, learn-plan, feynman-loop, learn-grade, learn-retain, learn-practice, learn-certify, learn-kb, learn-about-system, learn-harness), FSRS-6 spaced retrieval, KB adapters, and meta-learning for the Prometheus stack
10Language & Domain SkillsRust, React, Flutter, Tauri, HTMX, TypeScript, Go, Python, architecture, testing, DevOps, document extraction, and the Flint SDK skills
11The Artifact RefinerThe artifact-centric refinement engine and all fifteen of its commands
12The Native Agent GeneratorGenerating complete Rust agents, the A2A/AG-UI/A2UI protocols, and agent networks

The engine room — tools, toolchain, lifecycle​

#PageWhat it covers
13Tools Referenceforge-rs, prometheus-cli, prometheus-knowledge, liter-llm, surreal-memory-server, prometheus-rust-auditor, and prometheus-exec
14The Rust Toolchain & Dynamic GenerationWhy Rust, how the binaries are built, and how the pack generates new skills, CLIs, and MCP servers
15Hooks & LifecycleEvery hook event and script, progress signaling, scope guards, and the immutable-tests rule
16CLI & Scripts ReferenceEvery installer, validator, and runtime script, plus the npm script surface

Deployment — install, run, update, contribute​

#PageWhat it covers
17Platform SupportPer-tool support: Claude Code, OpenCode, Codex, Kimi Code, MiniMax, Cursor, Windsurf, Gemini CLI, Roo Code, Amp
18Plugins & MarketplaceClaude Code plugins and marketplace, OpenCode plugins, and the distribution model
19InstallationToolchain install (Rust, Go, Node, Docker), the one-command install, and MCP services
20UpdatingKeeping skills, tools, submodules, and MCP services current without breaking anything
21ContributingThe open-source workflow, validation gates, submodules, and importing skills

Closing​

#PageWhat it covers
22Advantages & ImpactWhat changes about your development process, and why
23Glossary & SourcesEvery term defined, every external claim cited

Operational​

DocumentWhat it covers
Production Readiness ReportEvidence table separating artifact, disposable-runtime, installed-service, and external-deployment certification
Deployment ModesMode 0-3 capability matrix — which services are required for which features
Dynamic Operations with Prometheus ExecChoosing generated programs versus native agents, Tier P/W/R theory, APIs, receipts, examples, security, and platform evidence

The thirty-second version​

If you read nothing else, read this.

A loop that forgets is a very fast way to do the same thing many times. A loop that remembers gets better at the task it was built to do. The prometheus-skill-pack is the infrastructure that makes the second kind of loop viable — and makes it work the same way no matter which AI tool you point at it.

The agents are ready. The substrate is the question. This is the answer to that question.


A note on accuracy​

This documentation is held to the same standard as the system it describes. Every external claim is cited in the Glossary & Sources. Narrative sections were checked against the sycophancy-correction MCP server before publication — the same structural quality gate the skill pack runs on its own reflection output. Where the repository contains a known inconsistency (for example, version drift across an imported skill's manifests), this documentation names it rather than papering over it. A document that only tells you what works is not documentation. It is marketing.