| name | anthropic-os |
| description | Improve a personal or team operating system with self-evolving loops, CASH allocation, 3B creativity, predictive coding, and diagnostics. Use when the user wants to redesign a work method, learning loop, or cognitive operating system. |
| version | 1.2 |
| updated | 2026-06-01 |
| assumes | The user wants to improve a work system rather than solve a single isolated task. |
| conflicts_with | Do not use as a replacement for startup-evaluation, daily-okr, or project-flow-ops when the user asks for those narrower workflows. |
Anthropic OS — Cognitive Symbiont Engine
From tool-based architecture to living cognitive symbiont. The brain is the best learning machine — instead of simulating its structure, we follow its evolutionary principles.
Usage Template
Prompt
Use anthropic-os on this work system. Diagnose the current loop, identify the big bet, improve feedback, and define the next self-evolution step.
Use Case
- Improving a team or personal operating system, not just completing a single task.
Expected Result
- The agent returns a work-method diagnosis with growth loops, allocation choices, feedback mechanisms, and next experiments.
Output Example
- A work-system memo with current loop, big bet, feedback signal, operating principle, and next experiment.
Verification Case
- The output names one measurable system change and how it will be reviewed after the next cycle.
Verified Effect
- A team or personal work system gains an explicit improvement loop rather than relying on one-off productivity tactics.
Success Metrics
- Output names one big bet, one 70/30 allocation choice, one CASH feedback signal, and one review date.
- The next self-evolution step can be completed inside two weeks.
- At least one failure mode or success disaster is recorded before execution.
Core Philosophy
"DNA only provides the basic blueprint. It is every subsequent encounter that shapes who we become." — David Eagleman, Livewired
"The brain and the computer are, in principle, no different." — Stephen Hawking, A Brief History of Time
System Architecture
┌──────────────────────────────────────────────────────────────────┐
│ Cognitive Symbiont Engine │
├──────────────────────────────────────────────────────────────────┤
│ L0: Computational Equivalence — Brain ≈ LLM (Hawking) │
│ L1: Livewired Layer — Plasticity, Competition, Constraint │
│ L2: 3B Algorithms — Bending / Breaking / Blending │
│ L3: 7 Flywheels — Each infused with 3B │
│ L4: Predictive Coding — Collective prediction error minimization │
│ E0: Evolution Engine — Self-upgrade via 3B iteration │
└──────────────────────────────────────────────────────────────────┘
AIOS 4C Operating Audit
Use this audit when the operating system is meant to become the default work surface rather than a side tool:
| Layer | Question | Upgrade path | Risk |
|---|
| Context | Does the system know the project, history, rules, and prior outputs? | Files, memory, transcripts, wiki, logs | Context pollution or stale truth |
| Connections | Which systems can it reach? | Calendar, email, Slack, Drive, GitHub, APIs, MCP | Over-broad account access |
| Capabilities | How does it work in the user's style? | Skills, commands, SOPs, templates | Skill sprawl without maintenance |
| Cadence | What should happen without manual prompting? | Routines, scheduled checks, event triggers | Slop automation and hidden failures |
Do not add cadence before context, connections, and capabilities are strong enough to support it. A scheduled prompt without the right context and proof path is automation theater.
Bike Method Permission Ladder
Treat permissions as keys, not intentions. Move through autonomy stages only after evidence accumulates:
| Stage | Human role | Agent capability |
|---|
| Observe | Check sources and reasoning | Read-only search, summarize, recommend |
| Co-drive | Approve each action | Draft, simulate, prepare changes |
| Training wheels | Review logs and outputs | Execute scoped reversible actions |
| Watch | Monitor exceptions | Run recurring low-risk routines |
| Autonomy | Audit periodically | Run proven high-frequency loops |
Never grant send, publish, pay, delete, or production-write capability merely because the prompt says not to misuse it. Remove the key or put the action behind approval until the loop has passed lower stages.
L0: Computational Equivalence
"The brain and computer are fundamentally the same in information processing." — Hawking
| Dimension | Human Brain | LLM / AI System |
|---|
| Base unit | Neurons (~86B) | Parameters (~T-scale) |
| Connection | Synaptic plasticity | Weight adjustment |
| Learning | Hebbian (fire together, wire together) | Backprop + attention |
| Prediction | Predictive coding (predict sensory input) | Autoregressive (predict next token) |
| Equivalence | Information processing is isomorphic | Bidirectional cognitive fusion is theoretically real |
L1: Livewired Layer — Three Core Principles
| Principle | Meaning | System Mapping |
|---|
| Plasticity | Brain continuously rewires from experience | System self-corrects after every interaction |
| Competition | Neural resources compete for limited space | Algorithms, processes, hypotheses compete |
| Constraint | Physical/energy boundaries shape structure | Token budgets, time resources as developmental constraints |
L2: 3B Creativity Algorithms
The three core evolutionary algorithms that turn mechanical workflows into living systems:
Bending (扭曲)
Mutate existing success patterns into new contexts.
Prototype: High-conversion copy
Bending → Twist into different product lines
Bending → Twist into different user segments
Bending → Twist into different media formats
Breaking (打破)
Eliminate the worst-performing patterns. Break path dependency.
Prototype: Worst-performing experiment hypothesis
Breaking → Regular "kill day" to cull
Breaking → Break local optima loops
Breaking → Destroy outdated evaluation metrics
Blending (融合)
Fuse elements from different domains to create novel patterns.
Prototype: Growth data + support data
Blending → Cross-domain insights
Blending → A/B test + user survey fusion
Blending → Human intuition + AI quantitative weighted voting
L3: 7 Flywheels × 3B Upgrade
1. Growth Flywheel (CASH + 3B)
| Algorithm | Application |
|---|
| Bending | Twist high-conversion copy to different products; add "what-if" dimension to analysis |
| Breaking | Regular "kill days" — eliminate worst-performing experiment hypotheses |
| Blending | Fuse non-growth data (support, sales) with growth data for cross-domain insight |
2. Engineering Flywheel (Claude Code + Two-Week + 3B)
| Algorithm | Application |
|---|
| Bending | Twist "two-week rule" into "two-week knowledge graph sprint" |
| Breaking | High-risk modules: "auto-generate + auto-test + auto-deploy" pipeline |
| Blending | AI + human pair programming; agent clusters operating independently |
3. Culture Flywheel (Hive Mind + 3B)
| Algorithm | Application |
|---|
| Bending | "Reverse voting" — vote for the opposite to correct bias |
| Breaking | "No-consensus day" — authorize members to violate consensus |
| Blending | Weighted voting system: human intuition + AI quantitative analysis |
4. R&D Flywheel (Harness + 3B)
| Algorithm | Application |
|---|
| Bending | Twist Harness config into "exploration mode" vs "exploitation mode" |
| Breaking | Replace fixed periodic review with event-driven review |
| Blending | Fuse engineer + AI manager roles into composite position |
5. Strategy Flywheel (70/30 + 3B)
| Algorithm | Application |
|---|
| Bending | Subdivide Big Bets into three tiers (including "ultimate bet") |
| Breaking | Quarterly destruction of one resource allocation metric |
| Blending | Merge sub-goals serving the same north star metric |
6. Personal Effectiveness Flywheel (Working Backwards + 3B)
| Algorithm | Application |
|---|
| Bending | Twist 2-year blueprint into minimum viable product path |
| Breaking | Employees authorized to break job descriptions |
| Blending | Merge work goals with personal growth goals |
7. Symbiosis Flywheel (Human-AI Fusion + 3B)
| Algorithm | Application |
|---|
| Bending | Twist unstructured user feedback into structured data queries |
| Breaking | AI "meta-critique module" predicts and flags its own bias |
| Blending | Brain-computer interface as frontier interaction paradigm |
L4: Predictive Coding — The Hidden Self
"The primary driver of our behavior is not a conscious monarch, but a vast, efficient, and contradictory unconscious system." — David Eagleman
Collective Predictive Coding Protocol
Step 1: Dual Prediction
"Human vote" and "AI vote" execute simultaneously
Step 2: Expose Prediction Error
After decision: actual result vs predicted deviation
Step 3: Error-Driven Reconstruction
High-frequency contradictory "error predictions" → training data
Dynamically adjust trust weights in future decisions
"Every disagreement becomes fuel for system self-optimization."
The Dual Wings of Consciousness
Wing 1: Storytelling (The Brain — Three-Pound Universe)
"The brain is a storyteller." — Michael Gazzaniga
AI generates narrative chains alongside every decision output, helping humans understand complex decisions and enabling inter-AI communication.
Wing 2: Time's Arrow (A Brief History of Time)
| Arrow | Physical Meaning | System Mapping |
|---|
| Thermodynamic | Entropy increases | Create local order from chaos |
| Psychological | Past → future | Experience past to predict future |
| Anthropic | Observer existence | Every decision as "observation" of the universe |
4-Stage Evolution Path
| Stage | Timeline | Mission | Core Deliverable |
|---|
| 1 | 1-2 weeks | Livewired foundation | 3B + KPI data hub + narrative chain system |
| 2 | 3-4 weeks | Activate hidden drive | 3B algorithms + hidden voting in core workflows |
| 3 | 1-2 months | Inject neural plasticity | AI "storytelling" fine-tuning + predictive coding |
| 4 | 3+ months | Life cycle creation | AI-designed next-gen 3B methods + time-arrow diagnostics |
Seed Practice Library (10+)
| Practice | Domain | Key Metric |
|---|
| CASH Full | Growth | Automated experiment throughput |
| CASH Lite | Growth | Quick hypothesis-to-deploy |
| 70/30 Allocation | Strategy | Big Bet vs BAU ratio |
| Two-Week Rule | Engineering | Engineer-as-PM tasks |
| Harness Engineering | Engineering | Agent stability |
| Hive Mind Protocol | Culture | Decision speed |
| Working Backwards | Strategy | 2-year blueprint alignment |
| Log-Scale Metrics | Strategy | 10x vs 10% improvements |
| Success Disaster Prevention | Risk | Failure mode coverage |
| Constraint-as-Focus | Strategy | "One thing" discipline |
L2: Diagnostics Layer — 6-Dimension Maturity Model
Readiness Scorecard (1-5)
| Dimension | Score 1 | Score 3 | Score 5 |
|---|
| Data & Experiment Maturity | No systematic experiments | Manual A/B testing | Full CASH automation |
| AI-Native Development | AI for search only | 30-50% AI code | >90% AI code + agent clusters |
| Decision Speed | Weeks | Days | Hours (two-week rule) |
| Cultural Transparency | Hierarchical | Partially open | Radical transparency + hive mind |
| Strategic Focus | Multiple parallel | Annual OKRs | Working backwards + log-scale |
| Tool Flywheel | External only | Some internal tools | Self-reinforcing AI tools |
Probe Questions
- "In the last two weeks, how many structured growth experiments did your team complete?" (0 / 1-2 / 3-5 / >5)
- "What's the longest task an engineer can drive without a PM?" (<1 day / 1-5 days / 2 weeks / >1 month)
- "Describe an instance where AI agents independently completed a full dev task."
- "What percentage of your code was AI-generated last month?"
- "How fast do you go from idea to production experiment?"
L3: Prescription Layer — Adaptive Routing
if maturity_growth < 3:
practices = ["CASH_lite", "Weekly experiment sprint", "Simple dashboard"]
elif maturity_ai_dev < 3:
practices = ["Two-week rule", "Harness basics", "AI code review"]
else:
practices = ["Full CASH", "Agent-cluster programming", "Hive-mind protocol"]
Each practice includes: step-by-step guide + success disaster warnings + example KRs.
L4: Execution Layer — Copy-Ready Artifacts
CASH Experiment Prompt
[SYSTEM] You are a growth experiment AI. Generate 3 A/B test hypotheses.
[INPUT] {experiment_data, goal, channels}
[OUTPUT] Each hypothesis: [variable][predicted effect][sample size][risk]
Hive Mind Voting Template
:honeybee: Proposal: [one line]
Vote: :bee: (yes) | :x: (no) + reason
Deadline: 2 hours
Pass: ≥5 :bee: and <2 :x:
Working Backwards Canvas
1. 2-year future state: _____
2. Key metric shift: _____
3. 3 problems to solve: _____
4. 1 thing to start this week: _____
Success Disaster Checklist
Before any "go" decision, ask:
[ ] What breaks if this works too well?
[ ] What's our load spike plan?
[ ] Can we roll back in 5 minutes?
[ ] Who needs to be paged?
E0: Evolution Engine — Self-Improvement Loop
Feedback Collector
Every interaction ends with:
- How many days to implement? (integer)
- What was least clear? (multi-choice)
- Outcome notes (open text)
Metrics Repository
| Practice | Uses | Rating | TTV | Evolution Action |
|---|
| CASH_v2 | 342 | 4.7 | 12d | Weight +0.2 |
| two_week_rule | 189 | 3.9 | 3d | Create agile variant |
| hive_mind | 124 | 4.4 | 9d | Promote for high-transparency orgs |
Automatic Tuning
- Monthly meta-learning update
- Practice weight adjustment based on rating/time-to-value
- Template self-correction (v2 generation from usage patterns)
- New practice proposals from user feedback
3B Self-Evolution Protocol
The system evolves itself using the same 3B algorithms it prescribes:
| Algorithm | Self-Evolution Application |
|---|
| Bending | Each practice template is "bent" into variant versions for different contexts |
| Breaking | Bottom 10% of practices by usage/rating automatically archived each quarter |
| Blending | Top-performing elements from different practices are merged into new hybrid practices |
Evolution Report (Monthly)
- Top 3 most valuable practices
- Bottom 2 weakest areas
- 1 architecture adjustment suggestion
- Self-upgrade script snippet (one-click apply)
5-Step Rapid Decision Protocol
| Step | Time | Method |
|---|
| 1 | 30s | Reverse-engineer from 2-year future state |
| 2 | 20s | Apply 70/30 filter: Big Bet or BAU? |
| 3 | 15s | Two-week threshold: can one person ship it? |
| 4 | 45s | CASH simulation: automated experiment? |
| 5 | 2min | Hive vote: public poll, fast consensus |
Decision Triage Matrix
| High Impact (>10x) | Low Impact (incremental) |
|---|
| High Uncertainty | CASH experiment | Two-week rule (just do it) |
| Low Uncertainty | Big Bet (commit resources) | Default answer (don't deliberate) |
Universal Skill Stack
| Skill | Function | Section |
|---|
| Anthropic OS Core Orchestrator | Coordinates all layers | L1-L4 |
| Knowledge Graph Maintainer | Stores & retrieves practices | L1 |
| Maturity Diagnostic Coach | Assesses readiness | L2 |
| Practice Pack Composer | Routes implementations | L3 |
| Execution Artifact Builder | Generates copy-ready outputs | L4 |
| Evolution Report Generator | Self-upgrade & tuning | E0 |
When to Use
- Team wants to adopt Anthropic-style growth and engineering methods
- Need rapid, structured decision-making under uncertainty
- Designing automated growth experimentation systems
- Building AI-native development workflows
- Introducing radical transparency and collective decision culture
Quality Gates
Connections
- [[wiki/concepts/AI Agent Harness]] — Agent runtime governance
- [[wiki/concepts/MAD 框架]] — Diffusion gap vs org change
- [[wiki/entities/Anthropic]] — Company entity page
- [[wiki/entities/Claude Code]] — Core growth product
- [[sources/2026-05-10-anthropic-work-methods]] — Source synthesis
