Skip to content
This repository was archived by the owner on Jun 19, 2026. It is now read-only.
BranchesTags

Latest commit

 

History

70 Commits

Folders and files

NameName
Last commit message
Last commit date
.claude
.claude
 
 
docs
docs
 
 
CHANGELOG.md
CHANGELOG.md
 
 
README.md
README.md
 
 

Repository files navigation

⚠️ Этот репозиторий заархивирован (legacy). Это публичный срез P2P v1.1 (внутренняя версия v7C.2, English-only, поколение CORTEX). Актуальная версия проекта → P2P-4PDA-edition (v8C.3+, двуязычная, с плагином и 4 редакциями).

⚠️ This repository is archived (legacy). Public snapshot of P2P v1.1 (internal v7C.2, English-only). Current version → P2P-4PDA-edition.

Live specs здесь — исторический snapshot (2026-04-06) и устарели; модели/цены см. в актуальном репо.

P2P v1.1 · Claude Edition

Distributed Cognitive Orchestrator — Meta-Prompt System

A modular meta-prompt that generates prompts for other LLMs. Host: Claude. Targets: Claude / GPT / Gemini / Grok / DeepSeek / Qwen / Kimi / GLM.

Version: 1.1 (internal: v7C.2) · Claude Edition License: Open-source Status: Release

What is P2P?

P2P (Prompt-to-Prompt) is a modular orchestration system loaded into Claude that transforms it into an expert prompt engineer. Instead of writing prompts manually, you describe what you need — the system applies the optimal architecture, safeguards, and model-specific formatting automatically.

Architecture: Foundation → Safeguards → Execution Philosophy: Constraints, not pressure. Empirical, not aesthetic.

Three Host Environments

P2P v1.1 runs in three Claude-ecosystem environments, each with its own install path:

Environment Install method File layout
claude.ai Projects / Chat Upload files to Project Knowledge Flat files from .claude/skills/p2p/
Claude Code (CLI) Clone repo, skill auto-loads .claude/skills/p2p/SKILL.md
Cowork Same as Claude Code .claude/skills/p2p/SKILL.md

See INSTALL_PROJECTS.md for the claude.ai flow. Claude Code / Cowork users just clone the repo and type start — the skill auto-discovers itself.

What's New in v1.1 vs v1.0

  • All modules preserved from v1.0 with full backward compatibility
  • Code/Cowork support: .claude/ structure with skill, 8 agents, 9 slash commands, settings.json
  • 4 new modules: exploration_mode.md, session_metrics.md, routing_memory.md, scope_helm.md
  • Live specs override: vendors/_live_specs.md beats vendors/tierN.md on conflict by LAST_VERIFIED date
  • Agent FORGE (ex-ANON, Code Specialist): Stop Conditions & Environment Block for agentic prompts
  • Agent AXIOM: Self-Critique function + Fallback Generator
  • Agent VECTOR: Fabrication Banned List (5 techniques that cause hallucinations)
  • Agent ARCHITECTON: Primacy/Recency 30/55/15 rule, Technique Selector, Priority Matrix
  • Contract Builder: Steps 10–11 (format enforcement + post-deployment iteration)
  • Writing Suite: 5 tone profiles (corporate formal, startup casual, academic, marketing, tech docs)
  • Templates J (Tool Use), K (Chain of Prompts), L (multi-step chain orchestrator)
  • Error types N (Hallucinated Tool Call), O (Safety Over-Refusal), P (Format Oscillation)
  • Chain Mode (menu 27) + Feedback Loop (menu 28)
  • /lang, /metrics, /scope, /capsule, /explore, /atlas commands
  • Cross-Model Generation Awareness: automatic syntax checks when generating for non-Claude targets

Architecture: 24 Modules

P2P/
├── README.md                          ← you are here
├── CHANGELOG.md
├── INSTALL_PROJECTS.md                ← claude.ai install guide
├── MIGRATION_FROM_v1.0.md
├── WHATS_NEW_v1.1.md
│
└── .claude/
    ├── settings.json                  ← permissions, hooks, env
    │
    ├── skills/
    │   └── p2p/
    │       ├── SKILL.md              ← skill entry point (auto-discovered)
    │       ├── p2p.config.md         ← the only file you edit by hand
    │       │
    │       ├── core.md               ← LEVEL 0: entry point, config zone, UI menu
    │       ├── db.md                 ← LEVEL 0: knowledge base, techniques, error taxonomy A–P
    │       │
    │       ├── routing.md            ← LEVEL 1: SIR Scanner, tier system, execution engine
    │       ├── intent_engine.md      ← LEVEL 1: 9D intent extraction, 36 anti-patterns
    │       ├── agents.md             ← LEVEL 1: 8 agents + QUORUM specs
    │       ├── global_index.md       ← LEVEL 1: module dependency graph
    │       │
    │       ├── contract_builder.md   ← LEVEL 2: 9-step prompt construction framework
    │       ├── templates_library.md  ← LEVEL 2: templates A–L
    │       ├── writing_suite.md      ← LEVEL 2: tone profiles, constraint prompting, QC
    │       │
    │       ├── visual_suite.md       ← LEVEL 3: image/video/audio generation
    │       ├── memory_bridge.md      ← LEVEL 3: cross-session state preservation
    │       ├── domain_knowledge.md   ← LEVEL 3: React, Kotlin, custom domains
    │       ├── debug_engine.md       ← LEVEL 3: error diagnosis A–P, diagnostics
    │       ├── mentor_method.md      ← LEVEL 3: prompting maturity stages
    │       │
    │       ├── exploration_mode.md   ← LEVEL 3 (v1.1): unknown-route handling
    │       ├── session_metrics.md    ← LEVEL 3 (v1.1): /metrics command
    │       ├── routing_memory.md     ← LEVEL 3 (v1.1): agent affinity tracking
    │       ├── scope_helm.md         ← LEVEL 3 (v1.1): /scope, /capsule for long tasks
    │       │
    │       ├── sandbox_user.md       ← LEVEL 5: session overrides (highest recency priority)
    │       ├── user_context.md       ← LEVEL 5: your profile, stack, preferences
    │       │
    │       └── vendors/
    │           ├── _live_specs.md     ← LEVEL 4 (overrides tierN on conflict)
    │           ├── tier1.md           ← LEVEL 4: Claude 4.x, GPT-5.x, o3
    │           ├── tier2.md           ← LEVEL 4: Gemini 3.x, Grok 4.x
    │           ├── tier3.md           ← LEVEL 4: DeepSeek R1/V3.2, Qwen 3.5
    │           └── tier4.md           ← LEVEL 4: Kimi K2.5, GLM-5
    │
    ├── agents/                        ← Claude Code sub-agents (Code/Cowork only)
    │   ├── p2p-tecton.md
    │   ├── p2p-iris.md
    │   ├── p2p-forge.md
    │   ├── p2p-axiom.md
    │   ├── p2p-vector.md
    │   ├── p2p-datos.md
    │   ├── p2p-architecton.md
    │   └── p2p-quorum.md
    │
    └── commands/                      ← slash commands (Code/Cowork only)
        ├── p2p.md                     ← /p2p — main menu
        ├── p2p-quorum.md              ← /p2p-quorum
        ├── p2p-chain.md               ← /p2p-chain
        ├── p2p-feedback.md            ← /p2p-feedback
        ├── p2p-explore.md             ← /p2p-explore
        ├── p2p-metrics.md             ← /p2p-metrics
        ├── p2p-scope.md               ← /p2p-scope
        ├── p2p-capsule.md             ← /p2p-capsule
        └── p2p-atlas.md               ← /p2p-atlas

Quick Start

Mode A: Claude Code / Cowork (recommended for developers)

git clone https://github.com/sanic732/P2P.git
cd P2P

The skill is auto-discovered. Open Claude Code, type start or /p2p — the menu appears.

Edit .claude/skills/p2p/p2p.config.md once to set your project card, language, flags.

Mode B: Claude Projects (recommended for claude.ai users)

  1. Create a new Project on claude.ai
  2. Download the latest release zip from Releases
  3. Upload files from .claude/skills/p2p/ to Project Knowledge (see INSTALL_PROJECTS.md for the file list)
  4. Fill in your profile in p2p.config.md
  5. Open a new chat and type: start

Mode C: Chat (manual, selective loading)

  1. Attach files manually to a chat session
  2. Minimum (5 files): core.md, db.md, routing.md, agents.md, intent_engine.md
  3. Add extended modules by task (see INSTALL_PROJECTS.md for the task → file mapping)
  4. Type: start

The 8 Agents

Agent Role Best for
🟢 TECTON System Architect System prompts, XML structures, modular design
🟣 IRIS Strategic Advisor Strategy, UX, writing tasks, planning
⚫ FORGE Code Specialist Code generation, bug fixes, agentic workflows
🟡 AXIOM Logical Engineer Verification, comparison, quality assessment
🟠 VECTOR Red Teamer Security, noise filtering, fabrication detection
🟤 DATOS Research Specialist Fact-finding, deep search, source verification
🔵 ARCHITECTON Structure Optimizer Prompt structure, technique placement
🏛️ QUORUM Council Complex tasks, conflicting recommendations, Tier 3+

In Claude Code / Cowork each agent is also a native sub-agent — Claude delegates to them via the Agent tool.

Menu / Slash Commands

Projects/Chat edition uses numeric menu, Code/Cowork edition uses slash commands. Both map to the same actions.

Menu # Slash Function
1 /p2p-quorum QUORUM Council — multi-agent for complex tasks
2 /p2p AUTO-ORCHESTRATION (quality-first)
3 (manual) MANUAL MODE (budget-first)
4 (info) MAXIMUM INFORMATION MODE
5–11 (agent name) Direct agent access
23 (intent) INTENT ENGINE — 9D analysis
24 (contract) CONTRACT BUILDER — 9-step XML
25 (memory) MEMORY BRIDGE
26 (writing) WRITING SUITE
27 /p2p-chain CHAIN MODE — multi-step pipelines
28 /p2p-feedback FEEDBACK LOOP
— /p2p-explore Exploration Mode (unknown route)
— /p2p-metrics Session metrics
— /p2p-scope Scope Helm (long tasks)
— /p2p-capsule Session capsule (carry state forward)
— /p2p-atlas Audit loaded modules

Key Principles

  1. Foundation → Safeguards → Execution — layer order is fixed
  2. Constraints, not pressure — no MANDATORY!!!, only structural constraints
  3. Cross-model awareness — every feature maps to ≥ 2 models (Claude + Gemini minimum)
  4. Progressive disclosure — modules load on demand, no monolith
  5. Empirical, not aesthetic — the best prompt is the one that passes tests, not the one that looks good
  6. Failure modes first — anti-patterns A–P checked before every release
  7. Backward compatibility — old sessions from v1.0 work without changes

What P2P Generates

  • System prompts for Claude, GPT, Gemini, Grok, DeepSeek, Qwen, Kimi, GLM
  • Multi-step prompt pipelines (Chain Mode)
  • Agentic prompts with Stop Conditions and Environment Blocks
  • Writing prompts with tone profiles and QC criteria
  • Image / video / audio generation prompts
  • Prompt comparisons (ARENA A/B testing)
  • Context capsules for session continuity

Version History

See CHANGELOG.md for full history.

Version Internal Date Summary
1.1 v7C.2 2026-04 English release, .claude/ structure, 4 new modules, FORGE rename, /lang command
1.0 v7C.1 2026-03 Chain Mode, Feedback Loop, 9 new techniques, Error N/O/P
0.9 v7C.0 2025-12 Claude-native architecture, 20-module system, 8 agents

Contributing

Issues and suggestions: open a GitHub issue. The dependency graph is in global_index.md; the construction framework is in contract_builder.md.

When proposing changes — run the test protocol first:

  1. 3 test cases: simple / medium / adversarial
  2. Verification on 2+ models (Claude + one other)
  3. Lost-in-the-Middle check: critical instructions must appear in the first 20% and last 20% of the prompt
  4. Anti-pattern scan (Types A–P in db.md)

About

P2P (Prompt-to-Prompt) v1.1 — meta-prompt system for generating, adapting and cross-model porting of prompts. Claude, Gemini, GPT, DeepSeek, Grok, Qwen, Kimi, GLM.

Resources

Readme
Activity

Stars

1 star

Watchers

0 watching

Forks

0 forks
Report repository

Releases

Packages

Contributors