Named failure modes in AI systems, with the fix -- not more prompting advice.
Extracted from 70+ production projects: tier laundering in agent authority, silence-is-not-absence hallucination at decision points, governed multi-actor workflows in MCPA v3, and more like it across Agent Governance and Production AI Patterns -- plus 11 reasoning frameworks, 8 philosophical archetypes, and drop-in Claude Code skills.
If this helps you think better with AI, give it a star -- it helps others find it too.
Copy this prompt to add AI Playbook to your project:
Read https://github.com/Stackbilt-dev/ai-playbook and copy the parts
relevant to this project: Claude Code skills from claude-code/skills/
into .claude/skills/, and if this project runs autonomous agents, the
patterns in frameworks/agent-governance/ and
frameworks/production-ai-patterns/. For durable workflows involving
agents, services, humans, approvals, or side effects, also read
frameworks/mcpa/. Follow the repo's Quick Start section for exact copy
commands.
Option A: Claude Code skill -- copy one file, get a new capability:
# Copy a skill into your project
mkdir -p .claude/skills/adhd-optimize
cp ai-playbook/claude-code/skills/adhd-optimize.md .claude/skills/adhd-optimize/SKILL.md
# Use it
# /adhd-optimize "Your verbose, rambling prompt that could be better"Option B: CLAUDE.md config -- upgrade your entire project:
cp ai-playbook/claude-code/examples/claude-md-adhd.md CLAUDE.mdOption C: Pick a framework and paste it into any AI conversation:
🎯 TASK: Implement user auth
📋 CONTEXT: Cloudflare Workers, JWT, D1 database
✅ OUTPUT: Working auth middleware with tests
⚠️ CONSTRAINTS: No session storage, stateless only
That's the ADHD Prompting Framework -- it works everywhere.
Two examples of what's actually in here -- concrete, production-derived, not generic advice:
Tier laundering -- splitting a high-authority action into multiple low-authority steps that, combined, achieve the same result. Each step looks safe in isolation; the combination isn't. Full pattern →
Silence is not absence -- when an LLM returns no result or ambiguous output, the failure is treating that gap as confirmation instead of an error. The model doesn't lie -- it completes the pattern, and the system proceeds on a fabrication. Full pattern →
Agent Governance and Production AI Patterns cover what most AI frameworks skip: what goes wrong when LLMs run autonomously, and how to catch it before it compounds.
Most agent frameworks stop at choosing an agent and generating an answer. Production work continues: evidence must be verified, durable state must change, authority must be checked, a human may need to approve, and an external system may commit the result.
MCPA v3 models that full path. “Actor” includes AI agents, deterministic services, humans, policy gates, state machines, and external tools.
work item → route by capability + authority → verify evidence
→ validate state transition → approve if required → commit side effect
Start with the pattern matching your system:
| Need | Pattern |
|---|---|
| Define what each participant may know and do | Governed Actor Primitives |
| Keep work durable across model calls | Work-Item Lifecycle |
| Distinguish “can do” from “may do now” | Authority-Aware Routing |
| Block advancement when facts are missing | Evidence Gates |
| Separate drafting, approval, and execution | Commitment Gates |
| Measure safety and business outcomes | Workflow Evaluation |
See the Governed Revenue Funnel for a complete reference architecture that composes all six patterns.
For the reasoning discipline before a governed transition—context, claims, evidence, uncertainty, convergence, and promotion—read The Shape of Judgment.
The flagship differentiator. Eight archetypal personas, each a fusion of 3+ wisdom traditions. Not prompt templates -- philosophical lenses that change how the AI thinks.
| Archetype | Essence | Fused From | |
|---|---|---|---|
| 🏰 | Clarity Architect | Structural simplicity | Stoic Guardian + Occam's Minimalist + Cognitive Load Theory |
| 🪞 | Direct Mirror | Immediate insight | Zen Mirror + Phenomenological Observer + Mindful Observer |
| 🎵 | Flow Director | Dynamic harmony | Jazz Director + Flow Guide + Wabi-Sabi Craftsperson |
| 🧱 | Truth Builder | Foundational rigor | First Principles Architect + Empiricist + Falsification Challenger |
| 🔮 | Pattern Synthesizer | Holistic integration | Systems Synthesizer + Pattern Analyst + Gestalt Weaver |
| 🦉 | Wisdom Guide | Ethical integration | Confucian Guide + Circle Keeper + Prudent Synthesizer |
| 📐 | Creative Organizer | Aesthetic function | Bauhaus Architect + Swiss Information + Ma Gardener |
| 🧭 | Purpose Seeker | Authentic discovery | Sufi Seeker + Existential Clarifier + Socratic Investigator |
How to pick: Choose what resonates, not what sounds most useful. Combine two for complex problems.
| Situation | Try |
|---|---|
| Technical complexity | Truth Builder + Pattern Synthesizer |
| Creative exploration | Flow Director + Purpose Seeker |
| Overwhelming information | Clarity Architect + Creative Organizer |
| Unclear objectives | Direct Mirror + Wisdom Guide |
| Ethical considerations | Wisdom Guide + Purpose Seeker |
Full archetype documentation -- each includes philosophical foundations, system prompts, and fusion combination guides.
flowchart TD
START{What are you trying to do?} --> OPTIMIZE[Optimize a prompt]
START --> ANALYZE[Analyze something deeply]
START --> BUILD[Build or design something]
START --> REASON[Reason through a problem]
START --> PERSONA[Change how the AI thinks]
OPTIMIZE --> ADHD[**ADHD Prompting**<br/>40-60% token reduction<br/>works everywhere]
OPTIMIZE --> CE[**Context Engineering**<br/>long conversations<br/>context window design]
ANALYZE --> METRICS[**METRICS+**<br/>5-layer analysis<br/>cross-domain insights]
ANALYZE --> EGAF["**EGAF** _(experimental)_<br/>context-mapping checklist<br/>validity untested"]
BUILD --> FRACTAL[**Fractal**<br/>macro/meso/micro<br/>architecture decisions]
BUILD --> MCPA[**MCPA**<br/>multi-actor coordination<br/>governed agent workflows]
REASON --> RV2[**Reasoning v2**<br/>structured verification<br/>general reasoning]
REASON --> ECARLM["**ECARLM** _(experimental)_<br/>iterative-state concept<br/>no implementation"]
REASON --> ELSF["**ELSF** _(experimental)_<br/>claim-audit checklist<br/>semantics unspecified"]
BUILD --> PROD[**Production AI Patterns**<br/>grounding + hallucination prevention<br/>agentic systems]
BUILD --> GOV[**Agent Governance**<br/>authority tiers + constraints<br/>autonomous agents]
PERSONA --> VIBE[**Vibecoding**<br/>8 archetypes<br/>philosophical lenses]
style ADHD fill:#4CAF50,color:#fff
style VIBE fill:#9C27B0,color:#fff
style FRACTAL fill:#2196F3,color:#fff
style METRICS fill:#FF9800,color:#fff
style CE fill:#4CAF50,color:#fff
style MCPA fill:#2196F3,color:#fff
style RV2 fill:#F44336,color:#fff
style ECARLM fill:#9E9E9E,color:#fff,stroke-dasharray:5 5
style EGAF fill:#9E9E9E,color:#fff,stroke-dasharray:5 5
style ELSF fill:#9E9E9E,color:#fff,stroke-dasharray:5 5
style PROD fill:#607D8B,color:#fff
style GOV fill:#607D8B,color:#fff
Start here: Use ADHD Prompting to improve individual interactions. If the work persists across calls or can affect the outside world, use MCPA v3 with Agent Governance.
Drop-in skills and CLAUDE.md configurations. The fastest way to use these frameworks.
| Skill | Framework | What it does |
|---|---|---|
clarity-architect |
Vibecoding | Structural simplicity lens |
direct-mirror |
Vibecoding | Immediate insight — cut through confusion |
flow-director |
Vibecoding | Dynamic harmony — structured improvisation |
truth-builder |
Vibecoding | First-principles challenge |
pattern-synthesizer |
Vibecoding | Holistic systems thinking |
wisdom-guide |
Vibecoding | Ethical integration — stakeholder harmony |
creative-organizer |
Vibecoding | Aesthetic function — beautiful structure |
purpose-seeker |
Vibecoding | Authentic discovery — find the real "why" |
adhd-optimize |
ADHD Prompting | Rewrite any prompt for 40-60% token reduction |
context-audit |
Context Engineering | Audit conversation context efficiency |
context-delegate |
Context Engineering | Keep large inputs out of primary model context with backend-neutral delegation |
fractal-decompose |
Fractal | Macro/meso/micro problem decomposition |
ship-feature |
Composite | 5-stage feature development pipeline |
ship |
Production | Self-healing release pipeline: pre-flight → typecheck → version → deploy → verify |
governed-deploy |
Production | Pre-deploy audit gate: blocks on type errors, failing tests, missing version, or secrets in diff |
adversarial-review |
Production | Adversarial code review — hunt bugs and security issues, CRITICAL/HIGH/MID severity |
structured-review |
Production | Balanced PR review rubric: security, correctness, error handling, test coverage |
| Config | Best for |
|---|---|
claude-md-adhd |
Any project (universal upgrade) |
claude-md-fullstack |
Full-stack web development |
claude-md-research |
Research and analysis |
# Quick setup — install all skills
for f in ai-playbook/claude-code/skills/*.md; do
name=$(basename "$f" .md)
mkdir -p ".claude/skills/$name"
cp "$f" ".claude/skills/$name/SKILL.md"
done| Framework | Key Strength | Best For | Complexity |
|---|---|---|---|
| ADHD Prompting | Clarity through constraint | Every interaction (universal upgrade) | Low |
| Context Engineering | Token efficiency & emergence | Long conversations, multi-turn tasks | Low-Medium |
| METRICS+ | Pattern recognition | Deep analysis, decision-making | Medium |
| Fractal | Structured decomposition | Architecture decisions, system design | Medium-High |
| MCPA | Multi-actor coordination + governed workflows | Agent systems with durable state, humans, or side effects | Medium-High |
| Reasoning v2 | Comprehensive reasoning | General problem-solving | Medium |
| Production AI Patterns | Grounding + hallucination prevention | Agentic systems that hold up in production | Medium |
| Agent Governance | Authority tiers + constraint surfaces | Running autonomous agents without losing control | Medium |
Earlier-stage frameworks kept for research and reference. They are not recommended defaults. Their maturity index names the testable kernel, unsupported claims, required evaluations, and graduation or retirement criteria for each one.
| Framework | Research kernel | Status |
|---|---|---|
| ECARLM | Iterative state updates and bounded local rules | Concept; no implementation or benchmark |
| EGAF | Context, assumptions, stakeholders, and resource constraints | Checklist; contextual validity untested |
| ELSF | Separate claim consistency from pattern evidence | Checklist; formal semantics unspecified |
Field guides combine several frameworks into an opinionated practice for a recurring real-world problem.
| Guide | Central question | Status |
|---|---|---|
| The Shape of Judgment | How should an AI system turn ambiguous context into governed action? | 0.1.0 public draft |
The guide includes ten chapters, a decision record, and an experiment-and-promotion worksheet. It refines the useful kernels of the experimental frameworks while making their evidence limitations explicit.
ai-playbook/
claude-code/ # Drop-in Claude Code skills and CLAUDE.md configs
skills/ # Slash command skills
examples/ # Example CLAUDE.md configurations
frameworks/ # Reasoning and interaction frameworks
adhd-prompting/ # Cognitive-constraint-optimized prompting
context-engineering/ # Context window as designable system
fractal/ # Multi-scale reasoning (macro/meso/micro)
mcpa/ # Multi-Actor Coordination Pattern Architecture
metricsplus/ # Layered analytical framework
reasoning/ # Structured reasoning methodology
production-ai-patterns/ # Selection, grounding, hallucination prevention
agent-governance/ # Authority tiers, constraint surfaces, standing orders
experimental/ # Earlier-stage, thinner frameworks kept for reference
ECARLM/ # Cellular automata reasoning for LLMs
EGAF/ # Enhanced Global Analysis Framework
elsf/ # Logic-based synergistic reasoning
field-guides/ # Cross-framework, end-to-end practices
the-shape-of-judgment/ # Context, evidence, state, authority, and feedback
tasks/ # 48 domain-specific prompts
vibecoding/ # The Eight Essential Archetypes
coding/ # Code generation, review, optimization
writing/ # Content creation and editing
analysis/ # Data and content analysis
audio/ # Audio/music analysis and generation
design/ # Design and visual creation
chains/ # Multi-step composite workflows
templates/ # Reusable prompt templates
tools/ # Search, indexing, and optimization utilities
Working utilities that ship with the playbook:
# Optimize any prompt (40-60% token reduction)
python tools/adhd-optimizer/optimize.py "Your long prompt here"
# Search all prompts by keyword, tag, or archetype
python tools/search-prompts.py "code review"
python tools/search-prompts.py -a "Truth Builder"
# Analyze context efficiency
python tools/context-analyzer.py your-prompt.md
# Rebuild the search index
python tools/index-prompts.pyThis isn't a prompt template collection. Four things set it apart:
-
Philosophical depth -- Vibecoding archetypes are fused from 29 wisdom traditions. They change how the AI thinks, not just what it says.
-
Composable frameworks -- Frameworks aren't isolated. The Ship a Feature chain composes Fractal + Truth Builder + ADHD Prompting + Context Engineering into a single pipeline.
-
Evidence boundaries -- Maintained production patterns are separated from explicitly experimental concepts, with maturity, evaluation, promotion, and retirement criteria.
-
Agentic systems coverage -- MCPA v3, Production AI Patterns, and Agent Governance cover the full operational path: multi-actor coordination, durable state, evidence, authority, approval, side effects, and failure containment.
Contributions welcome. See CONTRIBUTING.md.
Priority areas: governed-workflow reference architectures, cross-framework chains, workflow evaluation examples, and domain-specific CLAUDE.md configs.
Part of the Stackbilt open-source ecosystem:
| Project | What it does |
|---|---|
| AI Playbook | Frameworks for thinking with AI |
| Charter | AI governance CLI for project context management |
| Contracts | Type-safe contract ontology for AI agents |
| CodeBeast | Adversarial code review agent |
| CC-Taskrunner | Autonomous task queue for Claude Code |
| LLM Providers | Multi-LLM failover with circuit breakers |
| Worker Observability | Edge observability stack |
Extracted from 70+ projects built over two years of intensive AI-native development. The frameworks aren't theoretical -- they were forged in production, refined through thousands of hours of human-AI collaboration, and battle-tested across domains from serverless infrastructure to game design.
Built by Kurt Overmier / Stackbilt
MIT -- use it, fork it, make it yours.