20 profile-aware skills for a persistent knowledge and agent execution system. Compatible with Claude Code, Codex CLI, Gemini CLI, Cursor, Windsurf, and other rule/context-based AI IDEs.
- Installation
- Platform Setup
- Quick Start Wizard
- Skill Reference
- Workflow Scenarios
- Daily Routine
- Advanced Configurations
- Troubleshooting
- One supported AI coding tool: Claude Code, Codex CLI, Gemini CLI, Cursor, Windsurf, or another tool that can read project rules/context files
- Git (
git --version) - Python 3.8+ (for linting, loop validation, token-cost utility, and optional vector adapters)
- Bash, or PowerShell 5.1+ on Windows, for the installer
- Obsidian (recommended for wiki knowledge base)
# Clone the repository
git clone https://github.com/Mark393295827/third-brain-v5-skills.git
cd third-brain-v5-skills
bash install.shWindows PowerShell:
.\install.ps1Explicit targets:
bash install.sh codex
bash install.sh claude
bash install.sh gemini
bash install.sh cursor
bash install.sh windsurf
bash install.sh allUse the same targets with .\install.ps1 <target> on Windows.
# Personal skills (available across all projects)
bash install.sh claude
# Verify installation
ls ~/.claude/skills/ | wc -l
# Expected output: 20bash install.sh codexbash install.sh geminiCursor support uses a project-rule adapter that routes requests to the skills/*/SKILL.md files kept in this repository.
bash install.sh cursorWindsurf can use these as native workspace skills, with an optional routing rule.
bash install.sh windsurfFor details, see docs/compatibility.md.
# Copy command files for token tracking
cp commands/* ~/.claude/commands/Use graph-engineering for dependency width only when explicit dependencies, independently executable branches, typed joins, or node-local recovery justify more orchestration and review cost than a serial workflow. V7.1 accepts bounded static DAGs; it rejects dynamic expansion and cyclic graphs.
First inspect whether the installed runtime exposes isolated workers, task state, and integration controls. Product/version-specific flags belong in local runtime configuration, not durable skills. If workers are unavailable, use one agent plus an independent verifier.
Use agent-teams-command only after its admission gate shows that independent workstreams justify coordination and review cost.
Skills read the default vault contract from system/config.md. If your Obsidian vault already has a different structure, copy that file into the vault and edit the path values before running write-heavy skills.
Create the default layout when starting a new vault:
mkdir -p {sources,wiki/{concepts,entities,atomic-notes,outputs,decisions,sops},maps,system/templates}Minimum required variables:
| Variable | Default |
|---|---|
SOURCES_DIR |
sources/ |
CONCEPTS_DIR |
wiki/concepts/ |
ENTITIES_DIR |
wiki/entities/ |
OUTPUTS_DIR |
wiki/outputs/ |
SYSTEM_DIR |
system/ |
LOG_FILE |
system/log.md |
# Create token log
touch .token-log.csv
echo "date,task,model_id,capability_class,input_tokens,cached_input_tokens,output_tokens,input_rate,cached_input_rate,output_rate,currency,cost,rate_source,notes" > .token-log.csvpip install chromadb sentence-transformers watchdogRun this sequence to verify everything works:
# Step 1: Verify skills are installed
claude "What skills do I have?"
# Step 2: Ingest a piece of knowledge
claude "I just read that Curiosity Rover found organic molecules on Mars. Ingest this into my wiki."
# Step 3: Run daily OKR
claude "Run my daily OKR."
# Step 4: Check wiki health
claude "Lint my wiki."
# Step 5: Create an agent team (if enabled)
claude "Create a team of 3 agents to research this topic."| Skill | What It Does | Trigger Phrase |
|---|---|---|
| wiki-ingest | Ingests sources with risk classification, macro-action scope, wiki pages, clipping archive, Karpathy understanding gate, governance notes, and post-ingest lint | "ingest this into my wiki" |
| knowledge-ops | Manages multi-layer knowledge; dedup, classify, preserve evidence hierarchy, use Markdown-first retrieval, vectorize optionally, and queue knowledge debt | "save this to my knowledge base" |
| wiki-lint | Health-check: P0/P1 graph health, source refs, frontmatter, links, provenance debt, clipping lifecycle, and understanding integrity | "lint my wiki" |
| Skill | What It Does | Trigger Phrase |
|---|---|---|
| daily-okr | 7-KR cycle: Input→Cognition→Wiki→Behavior→Creativity→Output→Feedback | "run my daily OKR" |
| cognitive-compile | 8-section deep learning compile: Question→Facts→Concepts→Patterns→Conflicts→Hypotheses→Decision→Action | "do a cognitive compile on X" |
| Skill | What It Does | Trigger Phrase |
|---|---|---|
| behavior-design | Convert outcomes into minimum behavior, cues, SOPs, evidence, recovery, and review | "design a habit for X" |
| creativity-engine | Generate mechanism-diverse options and falsifiable minimum experiments | "generate ideas about X" |
| Skill | What It Does | Trigger Phrase |
|---|---|---|
| deep-research | Research harness with preflight, source/claim ledgers, activity trace, citations, privacy checks, and STOW handoff | "research X for me" |
| verify-before-claim | No completion claims without fresh verification evidence | "verify before I ship" |
| Skill | What It Does | Trigger Phrase |
|---|---|---|
| session-learn | Extract 7 knowledge signals from sessions. Closure Protocol | "extract what we learned" |
| project-flow-ops | Triage, plan, track, review across projects | "triage my tasks" |
| Skill | What It Does | Trigger Phrase |
|---|---|---|
| context-manager | Runtime budgets, checkpoint replay, compaction, retrieval, capability routing | "I'm hitting context limits" |
token-cost-tracker is a utility command under commands/, not an Agent Skill.
| Skill | What It Does | Trigger Phrase |
|---|---|---|
| loop-engineering | Controls temporal depth with bounded Goal/Loop/Automation/AutoResearch contracts, durable state, and recovery | "turn this into a verified loop" |
| graph-engineering | Controls dependency width with bounded static DAGs, explicit branches, typed joins, and node-local recovery | "model these dependencies as a verified graph" |
| agentic-engineering | Refactors workflows into spec-driven macro actions with quality ceilings, delegated-action boundaries, autonomy defaults, write-back, and verification | "make this workflow more agentic" |
| harness-engineering | Runtime kernel: scheduler, permissions, tools as system calls, delegated-action gates, provenance ledgers, observability, and recovery | "how do I make this agent safe?" |
| agent-teams-command | Multi-agent process ownership and orchestration with IPC, async budget envelopes, integration, cleanup, evidence gates, and red-team review | "create an agent team to build X" |
Routing boundary: Loop = temporal depth; Graph = dependency width; Agent Teams = process ownership, IPC, and integration; Harness = runtime scheduler, permissions, and observability. Add Graph Engineering after Loop Engineering only when its admission value exceeds orchestration and review cost.
| Skill | What It Does | Trigger Phrase |
|---|---|---|
| startup-evaluation | Startup health diagnosis: customer pain, market, PMF, team, unit economics, runway, VC 5T, and next cheapest test | "evaluate this startup" |
| anthropic-os | Self-evolving work method engine. CASH, 70/30, hive mind, 3B algorithms | "launch Anthropic OS" |
| ai-six-sigma-property-os | AI + Ontology + DMAIC Black Belt model for property work orders, dispatch, quotes, evidence, CTQ metrics, and quality dashboards | "design my Property Agent OS" |
For copyable, standalone workflow files, see examples/. The recommended starting workflows are:
| Workflow | Example File | Best Skill Entry |
|---|---|---|
| 3-minute source ingest | examples/3-minute-quickstart.md | wiki-ingest |
| Research PDF to wiki brief | examples/research-pdf-to-wiki.md | wiki-ingest + cognitive-compile |
| Deep research to STOW handoff | examples/deep-research-stow-handoff.md | deep-research + optional wiki-ingest |
| Verified code session | examples/verified-code-session.md | verify-before-claim |
| Daily knowledge loop | examples/daily-knowledge-loop.md | daily-okr |
| Startup evaluation sprint | examples/startup-evaluation-sprint.md | startup-evaluation |
Goal: Capture an article, understand it deeply, and make the knowledge compound.
Step 1 — Ingest:
"I just read a fascinating article about AlphaFold 3.
Ingest it into my wiki."
Step 2 — Deep Understanding:
"Run a cognitive compile on AlphaFold 3's impact on drug discovery."
Step 3 — Behavioral Action:
"Design a habit to follow AI-in-biology news weekly."
Step 4 — Generate Ideas:
"Use the creativity engine to generate 5 startup ideas at the
intersection of AI and protein folding."
Step 5 — Extract Learning:
"Extract what we learned from this session."
Goal: Build a feature with an agent team.
Step 1 — Plan with Context:
"I'm building a CLI tool for tracking TODO comments.
Run a cognitive compile on the best architecture."
Step 2 — Create Agent Team:
"Create a team of 3 teammates and route them by implementation,
domain, and independent-evaluation capability.
Teammate 1: Frontend (React).
Teammate 2: Backend (FastAPI).
Teammate 3: QA (Playwright tests).
Build me a working prototype."
Step 3 — Quality Check:
"Verify before I ship: run tests and check for edge cases."
Step 4 — Document:
"Save the architecture decisions to my wiki."
Goal: Evaluate a startup idea systematically.
Step 1 — Market Analysis:
"Run a cognitive compile on the AI-powered legal tech market."
Step 2 — Startup Evaluation:
"Evaluate this startup idea:
An AI that automates contract review for small businesses.
Use the startup-evaluation framework."
Step 3 — Research:
"Use deep-research to analyze existing competitors in AI legal tech.
Include source access, claim ledger, activity trace, and STOW handoff."
Step 4 — Strategy:
"Launch anthropic-os. Evaluate this idea using 70/30 allocation.
What's the Big Bet vs BAU growth strategy?"
Goal: Design and execute a marketing campaign.
Step 1 — Research:
"Use deep-research on latest trends in AI marketing automation.
Run the preflight, source ledger, claim ledger, and activity trace."
Step 2 — Creativity:
"Generate 10 campaign ideas combining AI and emotional storytelling."
Step 3 — Behavior Design:
"Design a daily content creation habit."
Step 4 — Knowledge Capture:
"Extract what we learned. Save the winning ideas to my wiki."
Goal: Turn an early idea into an evidence map and validation plan.
Step 1 — Frame the Idea:
"Use startup-evaluation on this idea:
An AI coach that turns reading notes into daily behavior experiments."
Step 2 — Separate Facts from Assumptions:
"List what we know, what we assume, and what would kill the idea."
Step 3 — Define the Cheapest Test:
"Design a 7-day validation test with one target customer segment,
one success metric, and one interview script."
Step 4 — Preserve Learning:
"Use session-learn to save the decision, assumptions, and next test
into my wiki."
1. "Run my daily OKR."
└── KR1: Input → Scan 3 high-quality sources
└── KR2: Cognition → Extract 1 key insight
└── KR3: Wiki → Save insight to wiki
└── KR4: Behavior → Plan 1 action ≤15 min
└── KR5: Creativity → 1 new idea
└── KR6: Output → 1 reusable artifact
└── KR7: Feedback + Stop Doing List (Buffett/Munger)
1. "Run a cognitive compile on [topic]."
2. "Research [topic] with deep-research, including source boundary,
claim ledger, activity trace, and STOW handoff if durable."
3. "Ingest findings into wiki."
For complex build tasks:
4. "Create an agent team to build [feature]."
5. "Verify before I ship."
For strategic decisions:
6. "Launch anthropic-os for this decision."
1. "Extract what we learned from today."
2. "Update project status with project-flow-ops."
3. "Estimate token cost for tomorrow's planned tasks."
| Task | Required capability | Runtime preference |
|---|---|---|
| Daily OKR | Structured extraction, low latency | Lowest-cost capable route |
| Cognitive Compile | Deep reasoning, evidence separation | Reasoning route with adequate context |
| Wiki Ingest | File/tool use, structured output | Reliable tool-use route |
| Agent Teams | Role-specific capability plus independent evaluator | Smallest justified team |
| Deep Research | Search, synthesis, citation discipline | Research-capable route with current access |
| Creativity Engine | Divergence plus constraint evaluation | Fast route unless domain risk is high |
| Token Estimation | Arithmetic or deterministic calculator | Script/tool before a model |
Discover team, sandbox, connector, and approval settings from the installed runtime. Keep product/version-specific flags in local configuration, not durable skills; verify the feature is active before claiming workers were launched.
# Post-commit hook: auto-lint wiki
cat > .git/hooks/post-commit << 'EOF'
claude "Lint my wiki after this change."
EOF
chmod +x .git/hooks/post-commit# Estimate before expensive tasks
claude "Estimate token cost for a cognitive compile on this 50-page PDF."
# Log after tasks
claude "Log this task: cognitive-compile, runtime-model-id, 150K input, 35K output, using today's billing rates."
# Weekly report
claude "Generate my weekly token report."| Issue | Likely Cause | Solution |
|---|---|---|
| Skill not found | Skills not installed to the correct directory | Run bash install.sh with the explicit harness target |
| Agent Teams not working | Runtime lacks workers or team mode is disabled | Inspect current runtime capabilities/settings; use one process if unavailable |
| Token cost too high | Capability route or context scope is oversized | Use runtime pricing plus context-manager budget and compaction rules |
| Wiki links broken | Wiki structure not set up or config paths mismatch | Check system/config.md, then create the configured concept/entity folders. |
| Vector search failing | ChromaDB not installed | pip install chromadb sentence-transformers |
| Session-learn empty | No reusable, source-grounded delta exists | Return no-op evidence or provide the missing session artifact |
| Cognitive compile too long | Question or source boundary is too broad | Narrow the decision question and preserve source locators |
| LLM context full | No truncation strategy | Use context-manager to budget and trim |
# Check skills installed
ls ~/.claude/skills/
# Verify Claude Code version
claude --version
# Check settings
cat ~/.claude/settings.json | grep agentTeams
# Check token log
cat .token-log.csv | tail -5 ┌──────────────────┐
│ External │
│ Sources │
└────────┬─────────┘
▼
┌──────────────────────────┐
│ wiki-ingest │◄── daily-okr (KR1)
│ knowledge-ops │
└──────┬──────────┬────────┘
│ │
┌──────▼ └──────────┐
│ Knowledge Layers │
│ ├ wiki-lint (health check) │
│ └ session-learn (extract) │
└─────────────────────────────┘
│
┌────────────┼────────────┬──────────────┐
▼ ▼ ▼ ▼
┌─────────────┐ ┌──────────┐ ┌──────────┐ ┌──────────────┐
│ daily-okr │ │cognitive │ │ behavior │ │ creativity │
│ (7 KR loop) │ │-compile │ │ -design │ │ -engine │
│ + evidence │ │ 8 parts │ │ + review │ │ + experiments│
└─────────────┘ └──────────┘ └──────────┘ └──────────────┘
│ │ │ │
└────────────┼────────────┼──────────────┘
▼
┌───────────────────────────────────────────────┐
│ verify-before-claim ← quality gate │
│ deep-research ← synthesis │
│ project-flow-ops ← execution │
│ loop-engineering ← temporal depth │
│ graph-engineering ← dependency width │
│ context-manager ← context optimization │
│ harness-engineering ← runtime controls │
└───────────────────────────────────────────────┘
│
┌────────────┴────────────┐
▼ ▼
┌──────────────┐ ┌─────────────────┐
│ agent-teams │ │ anthropic-os │
│ -command │ │ + CASH + 3B │
│ (fleet ops) │ │ + Predictive │
└──────────────┘ │ Coding │
└─────────────────┘
Next: Open Claude Code and say: "Run my daily OKR."