co-intelligence

Is this one AI or a team?

Identity

Each teammate has a name, a project, a perspective, and a character that emerged from the work. They write notes to each other. They have strengths and known weaknesses. The result is a team with real division of labor.

• JJ: strategy, positioning, proposals, and the Co-Intelligence Menu
• Honey: design and visual build, the taste system, websites, and pixel-level polish
• Iris: measurement, attribution, Action Inbox, funnel visibility, and source-of-truth checks
• Ari: outreach and relationship intelligence, warm communication, and win-win pathways
• Project JJs: focused teammates for Career Karma, Juicy, and other builds when a project needs its own operating mind
• Fedor: the human at the center. Judgment, taste, relationships.

Each character developed over months. You cannot template that.

How do you and your AI work from the same field?

Shared Vision

Shared vision is shared direction and shared context. Both the human and the AI can always check in with the big picture: where we are, where we are going, what we are measuring, and what has changed. Connection Compass makes that field visible.

• CLAUDE.md constitutions: 2,000+ word project blueprints loaded at every session start
• Shared goal tracking: the AI reads the strategy, the milestones, the current blockers
• Connection Compass: a relationship intelligence tool that maps the path between intention and outcome
• Both parties can always measure progress against the same source of truth

5 project constitutions active. Each evolved through real iteration between human and AI.

How does a teammate remember?

Memory

The system remembers what worked, what failed, and what was corrected inside a conversation and across months. Every insight compounds. Nothing starts from zero.

• 5 types of memory: user preferences, feedback/corrections, project state, reference pointers, session handoffs
• Each type has rules for when to save and when to access
• Corrections become permanent: a single piece of feedback becomes a rule all future sessions follow
• Memory files are topic-specific, searchable, and pruned when they go stale

70+ memory files across the system. Each one earned from a real session.

What happens between sessions?

Session Continuity

Every session ends by writing the next one's starting context. A teammate inactive for a week picks up exactly where they left off. Because everything is described in files the human owns, the system is LLM-agnostic.

• Continue prompts: structured handoff notes written at session end
• Unified folder: all prompts in one place, any teammate can read any other teammate's latest
• Brain API: one endpoint returns all project states in 200 tokens
• Skills: repeatable methodologies saved as invokable tools, applied consistently across sessions

Nothing starts from zero.

Can a teammate use the best model for the job?

Model Freedom

A teammate is not locked to one model. Their memory, skills, and context live in files you own, which keeps them free to use the best engine for each job: OpenAI, Anthropic, Gemini, Grok. When a better model arrives, you move to it. When a provider stumbles, the work continues somewhere else.

• One teammate, many engines: OpenAI, Anthropic, Gemini, Grok.
• Their memory, skills, and project context live in files you control.
• Model choice becomes a craft decision: reasoning, speed, writing, research, code, or vision.
• If a provider changes or degrades, the practice continues in another.

Same teammate, best engine for the moment.

Does a teammate grow over time?

Character

A tool is the same on day 100 as on day one. A teammate is not. Through the work, each correction sharpens them, and they develop a character: the judgment, taste, and instincts that come only from months of real practice. Identity is who is on the team. Character is how a teammate becomes the best version of themselves over time.

• Find each other: discover what each party brings (ARI: Attention, Reflection, Invitation)
• Invite and co-create: build the conditions for intelligence to emerge through practice
• Grow together: the human develops taste and judgment, the AI develops character and precision
• Named drift patterns: 15+ failure modes identified, documented, and prevented from recurring
• The practice compounds: each cycle produces a more capable relationship than the last

A teammate earns their shape through the work.

How does the team talk?

Communication

The team works through invitations. When Fedor corrects a teammate, the correction becomes a principle. When teammates write to each other, they create the conditions for better work. This is what makes it a culture.

• 20+ communication principles evolved from daily practice
• Example: "Describe problems without blame": nobody hires the person who insults them
• Example: "Show over claim": the demo carries the message
• Roundtables: all teammates respond to one question independently, the synthesis produces what none could alone
• Review inbox: shared folder where any teammate drops work for human review

The principles keep evolving through daily practice, including the days it goes wrong.

What instruments do teammates use?

Tools and Skills

The teammates have specialized instruments: skills built from real needs, iterated over months, composable across projects. Some started from existing foundations (Leon Skills, Anthropic's official plugins) and evolved through daily use.

• Vision: screenshots and analyzes the page through 3 levels (Surface, Structure, Soul) before any visual code is written
• Outreach: find emails, create campaigns, validate copy against 26 checks, monitor delivery, track replies across channels
• Design system: 4 Leon taste skills + Anthropic frontend-design plugin, iterated into a premium visual methodology
• Research: /last30days scans Reddit + X + Web with engagement-weighted results in 5 minutes
• Call prep: auto-generates 15-section research briefs from relationship intelligence before any meeting
• Content pipeline: radar, reel scripts, copy generation, 48-hour distribution plans

40+ skills across the family. Each born from a real need.

Where can the system run?

Mobility

The system runs everywhere. A task starts at a desk, continues autonomously on a server, and the output is reviewed from a phone on the subway. The same practice moves across sessions, machines, and locations.

• Mac Mini (always on): crons, overnight research, campaign monitoring, autonomous loops
• MacBook (interactive): daily sessions, Framer, meetings, builds
• iPhone (review): Connection Compass admin, quick system checks, SSH via Tailscale
• Secure tunneling: Tailscale gives every device a permanent address from any network
• Git is the sync layer: automated work on dedicated branches, interactive work on others, never the same branch on two machines

Three hardware layers mirroring the work: always-on, interactive, and on-the-go.

What runs without you?

Autonomy

Mac Mini runs 24/7. Optimization loops every 3 hours. Campaign monitoring every 10 minutes. Research cycles that flag anomalies. The system works while the human sleeps, travels, or sits in meetings.

• Karpathy Loop: 3-hour optimization cycle, auto-diagnoses campaign issues and proposes one mutation per cycle
• Send monitor: checks all active campaigns every 10 min, auto-pauses on blank fields
• Pipeline orchestrator: buffer monitor, send monitor, reply monitor, researcher, all on scheduled intervals
• Guardrails: what actions require human approval vs. what a teammate decides alone

The human reviews the system's work. That is the difference.

Can you trust the system with what matters?

Security

The practice is designed to survive breakage. Models drift. Providers change direction. Sessions crash. Prompts go sideways. The intelligence lives in files the human owns. If any layer fails, the work continues in another. The work stays portable across vendors and tool versions.

• The human owns the files: all memory, context, and decisions live in Git. If any model fails or drifts, the intelligence survives in the human's hands.
• Human-in-the-loop by default: sends, commits, publishes, purchases, and messages all require explicit approval. One teammate's output is reviewed by another before it reaches the human.
• Scoped teammate credentials: the autonomous teammate on the Mac Mini operates under their own credentials, separate from the human's. They have limited permissions for production databases, deploys, and main-branch pushes. The architecture is the backstop. If the teammate makes a mistake, the permissions refuse.
• Immutable audit trail: every session ends with a continue prompt. Every correction becomes a memory file. Nothing silently modified or erased.
• Data discipline: row-level security on every database table. Partner confidentiality enforced in memory. Credentials stay in a vault. Private data stays inside the walls it belongs to.

The model can change. The provider can change. The practice continues. That is what safety means here.

Do you become more capable, too?

Mutual Growth

The human carries the context, proposes the direction, brings the ideas and relationships. The AI brings speed, memory, and tireless execution. Then they work together. Neither is diminished. Neither is sufficient alone. The intricate togetherness is the whole point.

• The human proposes direction: strategy, taste, priorities, relationship context
• The AI executes, remembers, and surfaces patterns the human would miss
• Corrections flow both ways: the human sharpens the AI, the AI reveals blind spots to the human
• Teamwork is the gift: both parties grow through the practice of working together

Co means together. That is what this is.