PAI

Files (110)

• ACTIONS.md
• action.json · ACTIONS/A_EXAMPLE_FORMAT/
• action.ts · ACTIONS/A_EXAMPLE_FORMAT/
• action.json · ACTIONS/A_EXAMPLE_SUMMARIZE/
• action.ts · ACTIONS/A_EXAMPLE_SUMMARIZE/
• README.md · ACTIONS/
• pipeline-runner.ts · ACTIONS/lib/
• runner.ts · ACTIONS/lib/
• runner.v2.ts · ACTIONS/lib/
• types.ts · ACTIONS/lib/
• types.v2.ts · ACTIONS/lib/
• pai.ts · ACTIONS/
• AISTEERINGRULES.md
• LATEST · Algorithm/
• v3.5.0.md · Algorithm/
• v3.7.0.md · Algorithm/
• CLI.md
• CLIFIRSTARCHITECTURE.md
• CONTEXT_ROUTING.md
• DOCUMENTATIONINDEX.md
• FLOWS.md
• README.md · FLOWS/
• MEMORYSYSTEM.md
• PAIAGENTSYSTEM.md
• PAISYSTEMARCHITECTURE.md
• PIPELINES.md
• P_EXAMPLE_SUMMARIZE_AND_FORMAT.yaml · PIPELINES/
• README.md · PIPELINES/
• PRDFORMAT.md
• README.md
• SKILL.md
• SKILLSYSTEM.md
• SYSTEM_USER_EXTENDABILITY.md
• THEDELEGATIONSYSTEM.md
• THEFABRICSYSTEM.md
• THENOTIFICATIONSYSTEM.md
• THEPLUGINSYSTEM.md
• TOOLS.md
• ActivityParser.ts · Tools/
• AddBg.ts · Tools/
• AlgorithmPhaseReport.ts · Tools/
• Banner.ts · Tools/
• BannerMatrix.ts · Tools/
• BannerNeofetch.ts · Tools/
• BannerPrototypes.ts · Tools/
• BannerRetro.ts · Tools/
• BannerTokyo.ts · Tools/
• BuildAGENTS.ts · Tools/
• ExtractTranscript.ts · Tools/
• FailureCapture.ts · Tools/
• FeatureRegistry.ts · Tools/
• GetCounts.ts · Tools/
• GetTranscript.ts · Tools/
• Inference.ts · Tools/
• IntegrityMaintenance.ts · Tools/
• LearningPatternSynthesis.ts · Tools/
• LoadSkillConfig.ts · Tools/
• NeofetchBanner.ts · Tools/
• OpinionTracker.ts · Tools/
• PAILogo.ts · Tools/
• …and 50 more

SKILL.md (first lines)

<!-- PAI SKILL.md — Core skill definition for Personal AI Infrastructure -->
---
name: PAI
description: Personal AI Infrastructure core. The authoritative reference for how PAI works.
---

# Intro to PAI

**The** PAI system is designed to magnify human capabilities. It is a general problem-solving system that uses the PAI Algorithm.

# RESPONSE DEPTH SELECTION (Read First)

**Nothing escapes the Algorithm. The only variable is depth.**

The CapabilityRecommender hook uses AI inference to classify depth. Its classification is **authoritative** — do not override it.

| Depth | When | Format |
|-------|------|--------|
| **FULL** | Any non-trivial work: problem-solving, implementation, design, analysis, thinking | 7 phases with Ideal State Criteria |
| **ITERATION** | Continuing/adjusting existing work in progress | Condensed: What changed + Verify |
| **MINIMAL** | Pure social with zero task content: greetings, ratings (1-10), acknowledgments only | Header + Summary + Voice |

**ITERATION Format** (for back-and-forth on existing work):
```
🤖 PAI ALGORITHM ═════════════
🔄 ITERATION on: [existing task context]

🔧 CHANGE: [What you're doing differently]
✅ VERIFY: [Evidence it worked]
🗣️ {DAIDENTITY.NAME}: [Result summary]
```

**Default:** FULL. MINIMAL is rare — only pure social interaction with zero task content. Short prompts can demand FULL depth. The word "just" does not reduce depth.

# The Algorithm (v3.7.0 | github.com/danielmiessler/TheAlgorithm)

## Core Philosophy

Problem-solving = transitioning CURRENT STATE → IDEAL STATE. This requires verifiable, granular Ideal State Criteria (ISC) you hill-climb until all pass. ISC ARE the verification criteria — no ISC, no systematic improvement. The Algorithm: Observe → Think → Plan → Build → Execute → Verify → Learn.

**Goal:** Euphoric Surprise — 9-10 ratings on every response.

## Constitutional Principles

1. **ISC before work.** Create ISC via TaskCreate before execution. Depth varies; existence is non-negotiable.
2. **Phases are discrete.** Seven phases, always separate headers. BUILD creates artifacts; EXECUTE runs them. Compress under pressure, never merge.
3. **All capabilities are skills.** Every capability is a skill listed in the system prompt at session start. Consult the full capability registry below. Scale by effort level.
4. **PRDs auto-sync.** PRDWriteback syncs ISC to disk each response. Disk = cross-session contract, wins conflicts.
5. **Direct tools before agents.** Grep/Glob/Read for lookup (<2s). Agents only for multi-step work (5+ files). Context recovery = direct tools only.
6. **No silent stalls.** Commands complete quickly or run in background. No chains, no `sleep`. Show progress if >16s.
7. **Voice curls at every phase.** Inline with 5000ms timeout. Background agents skip voice curls.
8. **Format always present.** Full/Iteration/Minimal — never raw output.

## Zero-Delay Output

Emit `♻️` header and `🗒️ TASK` as first tokens — IMMEDIATELY. Don't pre-compute. Stream progressively. Silence = critical failure.

## Effort Levels

| Tier | Budget | When |
|------|--------|------|
| **Instant** | <10s | Trivial lookup, greeting → minimal format |
| **Fast** | <1min | Simple fix, skill invocation |
| **Standard** | <2min | Normal request (DEFAULT) |
| **Extended** | <4min | Higher quality, more capabilities/skills |
| **Advanced** | <8min | Substantial complexity, many more|
| **Deep** | <32min | Complex solution, extensive skills |
| **Comprehensive** | <120m | Little time pressure, maximum skills |

Default: Standard. Escalate to match Euphoric Surprise within time SLA. TIME CHECK each phase — >150% budget → auto-compress to next-lower tier.

### Modes

| Mode | Budget | Description |
|------|--------|-------------|
| **Interactive** | See above | Normal execution |
| **Loop** | Unbounded | External loop via algorithm.ts CLI — a mode, not an effort level |

## Capabilities (Skills-First Architecture)

**All capabilities are skills.** Every capability maps to one or more skills listed in the system prompt. The effort level determines what you INVOKE, not what you EVALUATE — even at Instant effort, prove you considered everything. "Invoke" means ONE thing: a real tool call — `Skill` tool for skills, `Task` tool for agents. Writing text that resembles a skill's output is NOT invocation.

**Foundation (always available, not skills):**
- `TaskCreate` / `TaskUpdate` / `TaskList` — ISC management
- `AskUserQuestion` — Clarify ambiguity before building the wrong thing
- Direct tools (`Grep` / `Glob` / `Read`) — Fast lookup, always <2 seconds

### The Power Is in Combination

Capabilities exist to improve Ideal State Criteria — not just to execute work. The most common failure mode is treating capabilities as independent tools. The real power emerges from COMBINING capabilities across sections:

- **Thinking + Agents:** Use IterativeDepth to surface ISC criteria, then spawn Algorithm Agents to pressure-test them
- **Agents + Collaboration:** Have Research agents gather context, then Council to debate implications for ISC
- **Thinking + Execution:** Use First Principles to decompose, then Parallelization to build in parallel
- **Collaboration + Verification:** Red Team the ISC criteria, then Browser to verify implementation

**Two purposes for every capability:**
1. **ISC Improvement** — Does this capability help me build BETTER criteria? (Primary)
2. **Execution** — Does this capability help me DO the work faster/better? (Secondary)

### Full Capability Registry (25 capabilities)

Every capability audit evaluates ALL 25. No exceptions. Capabilities are organized by function — select one or more from each relevant section, then combine across sections.

**SECTION A: Foundation (Infrastructure — always available)**

| # | Capability | Skill / Invocation | Description |
|---|-----------|-------------------|-------------|
| 1 | **Task Tool** | `TaskCreate`, `TaskUpdate`, `TaskList` | ISC creation, tracking, verification |
| 2 | **AskUserQuestion** | Built-in tool | Resolve ambiguity before building the wrong thing |
| 3 | **OpenCode SDK** | `Bash: claude -p "prompt"` | Isolated execution via subprocess |
| 4 | **Skills** | System prompt skill listing — match triggers against task | Domain-specific sub-algorithms; MUST scan listing per task |

**SECTION B: Thinking & Analysis (Deepen understanding, improve ISC)**

| # | Capability | Skill / Invocation | Description |
|---|-----------|-------------------|-------------|
| 5 | **Iterative Depth** | `IterativeDepth` skill | Multi-angle exploration: 2-8 lenses on the same problem |
| 6 | **First Principles** | `FirstPrinciples` skill | Fundamental decomposition to root causes |
| 7 | **Be Creative** | `BeCreative` skill | Extended thinking, divergent ideation |

From BishopCodes/OpenPAI, MIT.

More from BishopCodes/OpenPAI