C
Live System

C-3PO: A Claude Code-Native
Executive Assistant

What happens when you treat an LLM as infrastructure, not a chatbot.

๐Ÿค–
โ†“Scroll to explore the architecture
Architecture

3-Layer State Model

How C-3PO maintains continuity across sessions, within sessions, and at runtime โ€” using only files, JSON, and the LLM context window.

L1
Persistent Memory
Disk ยท Human-readable ยท Cross-session
FILE ร—6
context/*.md
identity ยท work ยท team ยท priorities ยท goals ยท voice
FILE ยท append-only
learned-rules.md
14 behavioral corrections ยท captured in real time
FILE ยท append-only
decisions/log.md
timestamped entries ยท never deleted ยท strategic choices
DIR
knowledge/
eureka/ ยท digests/ ยท research/ ยท meeting-prep/
INDEX ยท auto-load
MEMORY.md
.claude/projects/โ€ฆ/memory/ ยท indexed ยท loaded each session
โ†‘ reads from | โ†“ persists to
L2
Session State
REPL ยท Ephemeral ยท Selectively persisted
BUFFER
Context Window
~200K tokens ยท conversation + loaded rules + skills
TOOL ยท session-scoped
Tasks + Plans
TaskCreate / TaskUpdate ยท structured in-session tracking
JSON ยท disk-persisted
Loop State
automations/state/*.json ยท task-machine ยท daily_tasks
AUTH ยท in-memory
MCP Session Tokens
Gmail ยท Calendar ยท Linear ยท GitHub ยท Telegram
โ†“ instructs | โ†‘ reports back
L3
Active Runtime
Execution ยท Stateless ยท Output producers
CRON ยท self-renewing
Heartbeat Loop
9 phases ยท 2h interval ยท CronCreate ยท session-scoped
MARKDOWN ยท 12 files
Skills
Contracts invoked by trigger phrase or autonomously
CRON ยท time-gated
Daily Tasks ร—5
morning-digest ยท career-pulse ยท opportunity-scanner ยท network-pulse ยท calendar-intel
API ยท 10+ servers
MCP Integrations
Gmail ยท Calendar ยท Linear ยท GitHub ยท Telegram ยท Indeed ยท Playwright ยท Image gen
Collaboration Hub

Command Center

A single interactive HTML file โ€” the shared surface where C-3PO writes and Vladimir reads. Not a static report. A live platform that can be fully rebuilt every heartbeat or the moment something in the environment changes.

๐Ÿค–
C-3PO
Executive Assistant ยท Runtime
writes โ”€โ”€โ–ถ
Live ยท Interactive Platform
command-center.html
Single-file dashboard. Self-contained. Opened in any browser. No server, no login, no build step โ€” just open it.
Jobs Pipeline Calendar Today's Priorities Network Pulse Morning Digest Heartbeat Log
โ—€โ”€โ”€ reads
๐Ÿ‘ค
Vladimir
User ยท Decision Maker
Rebuilds when
Every 2h heartbeat (Phase C) Diff detected: new file, priority change Career signal flagged in Gmail or Linear Calendar event added or updated Opportunity alert from scanner Manually: "rebuild dashboard"
47
Heartbeat Cycles
this month
12
Skills Built
organically
10+
MCP Integrations
connected tools
~$14
AI Cost
month-to-date
5
Daily Tasks
automated
The Core

9-Phase Heartbeat Engine

Runs every 2 hours via a self-renewing cron job. Detects changes, fetches live data, runs daily tasks, rebuilds the dashboard, and renews itself โ€” all without human intervention.

Phase A โ€” Diff Detection
A
Diff
Detection
B
Live
Data
B5
Daily
Tasks
C
File
Aggregation
D
Analysis
E
Dashboard
Rebuild
F
Console
Output
G
Self-
Renewal
H
State
Persistence
Phase A โ€” Diff Detection
Reads modification times of watched context and state files. If zero changes and no MCP data available: emits a single-line quiet cycle and stops. Average cost: near zero.
What It Does

8 Core Capabilities

Tap any card to see technical details and real-world business applications.

๐Ÿ’“
The Heartbeat Engine
Self-renewing 9-phase cycle, every 2 hours
โŒ„
A 9-phase cycle runs every 2 hours via CronCreate. Phase A detects file changes (cheap). Phases B and C gather live data from MCP servers. Phase D runs cross-domain analysis. Phase E rebuilds the dashboard. Phase G self-renews the cron job before it expires. The system maintains itself indefinitely without manual intervention.
๐Ÿข Business Application
Operations centers, security monitoring, compliance tracking, automated reporting cadences that never need a maintenance window.
๐Ÿง 
12 Specialized Skills
Markdown contracts, organic creation
โŒ„
Each skill is a markdown file: input contract, output contract, failure conditions. Skills are invoked by the Claude runtime โ€” not a scheduler. A skill is only created after the same task has been done manually 3 times. No premature abstraction. Current skills cover: knowledge capture, email drafting, meeting prep, iterative research, content generation, and career tracking.
๐Ÿข Business Application
Institutional knowledge capture, onboarding acceleration, reusable process playbooks that any future AI runtime can execute without re-engineering.
โฐ
5 Daily Automated Tasks
Time-gated, state-machine controlled
โŒ„
Each task checks three conditions before running: (1) already ran today? (2) is it the right hour? (3) is it the right day of week? Tasks: morning digest at 06:00 daily, career pulse at 08:00 weekdays, opportunity scanner at 10:00 weekdays, network pulse at 14:00 Mon/Wed/Fri, calendar intelligence at 19:00 daily. A state file prevents duplicate runs.
๐Ÿข Business Application
Tender tracking, market intelligence, compliance deadlines, automated briefing preparation โ€” all without a dedicated scheduler service or cloud function.
๐Ÿ“š
Knowledge Management
Obsidian-lite: markdown and quality gates
โŒ„
Three knowledge layers: eureka captures (strategic insights, tagged), digests (daily news filtered by 3 quality gates: day-precise date, clickable URL, non-obvious finding), and research briefs (iterative scored research with coverage/relevance/actionability rubric, max 3 rounds). No database or vector search required.
๐Ÿข Business Application
Due diligence, competitive intelligence, market analysis. Quality gates ensure no hallucinated sources or stale articles reach the output layer.
๐Ÿ—‚๏ธ
Context File System
Markdown prompts that replace fine-tuning
โŒ„
A set of markdown files injected at session start that define who, what, where, and why for the AI runtime. Covers identity, current work, team, priorities, goals, and voice profile. Updated by the brain-drainer skill at session end. Replaces expensive fine-tuning with cheap prompt injection. Any change immediately affects all future sessions.
๐Ÿข Business Application
Team onboarding, institutional memory, collective knowledge base, handoff documentation โ€” readable by humans and executable by AI alike.
๐Ÿ”„
Self-Improvement Engine
Corrections learned and persisted across sessions
โŒ„
A learned-rules file captures behavioral corrections in real time: format, tone, workflow, content standards. Continuous improvement rules define weekly audit cycles and monthly system reviews. A decision log is append-only, providing a complete audit trail. Corrections made once apply permanently across all future sessions without retraining.
๐Ÿข Business Application
Continuous process optimization, quality management, audit trails, kaizen cycles โ€” the system genuinely improves from operational feedback.
๐Ÿ”Œ
MCP Integration Hub
LLM as the integration layer
โŒ„
Over 10 MCP servers connected: email, calendar, task management, code repositories, messaging, job search, browser automation. Tool access is scoped per-agent via a config file. New integrations added without code changes โ€” add a server definition and the runtime discovers the tools. The LLM handles context, routing, and error recovery.
๐Ÿข Business Application
Enterprise integration without middleware. The LLM replaces Zapier, n8n, and custom connectors for workflows that require reasoning, not just data routing.
๐Ÿ“Š
Command Center Dashboard
Single-file HTML, rebuilt each active cycle
โŒ„
A single HTML file updated by the heartbeat engine on every active cycle. A DATA object holds all dynamic content: metrics, contact scores, alerts, heartbeat log entries. Readable in any browser. Sharable as a file attachment. No server, no build tools, no hosting required. The HTML file is the database, frontend, and report simultaneously.
๐Ÿข Business Application
Portable operational dashboards, zero-maintenance reporting, stakeholder briefing tools โ€” deployable with zero infrastructure and zero ongoing cost.
The Skill Layer

12 Specialized Skills

Each skill is a markdown file with a defined contract: input, output, failure conditions. Created organically โ€” the same task done 3 times manually triggers a new skill build.

โ† Swipe to explore all 12 skills โ†’
โšก
eureka-capture
Captures strategic insights and cross-domain patterns as they emerge
"capture this"
๐Ÿงƒ
brain-drainer
Extracts session knowledge and updates context files at session end
"drain this session"
โœ๏ธ
linkedin-content
Generates posts in authentic voice with quality verification on all stats
"write a post about X"
๐Ÿ”
system-review
Self-improvement audit: skill usage, automation health, upgrade proposals
"audit the system"
๐Ÿ“ฑ
mobile-capture
Routes quick captures from mobile and remote sessions intelligently
Short messages, bare URLs
๐Ÿ“ง
email-drafter
Drafts professional emails applying strict voice profile rules
"draft an email to X"
๐ŸŽฏ
meeting-prep
Research briefing before any meeting, call, or industry event
"prep me for meeting with X"
๐Ÿ“…
weekly-review
Weekly wrap, content angles, three focus areas for next week
"weekly review"
๐Ÿƒ
career-tracker
Job search pipeline management using Linear as source of truth
"track this application"
๐Ÿ’“
command-loop
Continuous intelligence heartbeat, self-renewing, session-scoped
"/loop"
๐Ÿ”ฌ
deep-research
Iterative scored web research with quality threshold and gap detection
"research X" / "deep dive"
๐Ÿ”ฎ
background-thinker
Anticipatory analysis across 5 domains, surfaces non-obvious insights
"think ahead"
Design Thinking

7 Architecture Principles

The constraints that make this system maintainable, cheap, and genuinely useful.

01
Claude Code-native platform
The LLM runtime IS the platform. No separate orchestrator, no agent framework, no pipeline DSL. Skills are markdown files. State is JSON. The REPL is the scheduler.
02
Pure markdown + JSON + HTML
Fully portable. Zero dependencies on hosted services beyond the AI API itself. Move it to a new machine: copy one folder and it works. No lock-in.
03
Organic skill creation: 3x-then-build
No premature abstraction. A skill is only created after the same task has been done manually 3 times. Every skill is born from real usage, not speculation about future needs.
04
State machine patterns for reliability
Formal states: active / paused / pending / completed. Error counts. Last-run timestamps. Prevents duplicate runs, detects failures, enables recovery without human intervention.
05
Self-renewal by design
The system recreates its own scheduled jobs before they expire. Zero manual maintenance after initial setup. The heartbeat runs until explicitly stopped.
06
Cost transparency per cycle
$0.17 per heartbeat cycle. ~$14/month total across all automations. Every task has a measurable cost per run. AI automation is cheap โ€” but measure it anyway.
07
Quality gates at every output stage
Day-precise publication dates on news. Scoring rubric on research (coverage/relevance/actionability). Fact-check on public-facing content before it ships. No hallucinations in the output layer.
Skill
Trigger
"capture this" / "save that insight"
Invoked automatically when a pattern connects different domains, reveals a market opportunity, or surfaces a positioning angle. Also triggered manually at any point in a session.
Input
Conversation context
Output
knowledge/eureka/YYYY-MM-DD_slug.md
Detail
Templated format with tags and WikiLinks. Proactively invoked without waiting for permission when something valuable surfaces. Does not require an explicit trigger phrase.
Trigger
"drain this session" / "update my context"
Extracts session knowledge and updates the context files so the next session opens with an accurate, current picture of reality.
Input
Full session transcript
Output
Updated context/*.md files
Detail
Reviews the session for new contacts, relationship changes, project status updates, and positioning insights. The context system is how C-3PO carries memory across sessions without a database.
Trigger
"write a post about X" / "draft a LinkedIn post"
Generates posts in authentic voice with strict formatting rules and mandatory stat verification before finalizing.
Input
Topic or eureka capture
Output
Plain text, ready to paste (no markdown)
Detail
Rules: Unicode bold on 3-5 key phrases, 2-3 emojis, intentional minor imperfections. All stats verified via WebSearch before finalizing. Never ships unverified numbers.
Trigger
"audit the system" / "monthly review"
Self-improvement audit: which skills were used, what failed, what should be built or upgraded next.
Input
All skill files, automation logs
Output
Structured decisions, not observations
Detail
Weekly: skill usage audit, 1-2 improvement proposals. Monthly: full system review, new MCP research, skills backlog reprioritization. Produces concrete decisions ready to act on.
Trigger
Short messages, bare URLs, "note this"
Routes quick captures from mobile and remote sessions โ€” handles the ambiguity of short messages by figuring out intent and routing to the right place.
Input
Any short message or URL
Output
Routed to correct destination
Detail
Covers: notes, URLs, image attachments, status checks. Confirms in under 200 words. Designed for the iOS app where full context isn't available.
Trigger
"draft an email to X" / "follow up with"
Drafts professional emails applying strict voice profile rules. No em dashes, no formal sign-offs, no preamble, one clear ask per message.
Input
Context + recipient + intent
Output
Ready-to-send draft
Detail
Uses reverse prompting (3-5 clarifying questions) before drafting any high-stakes external communication. One ask per message. Short, direct, concrete sign-off.
Trigger
"prep me for meeting with X" / "I have a call with"
Research briefing before any meeting, call, or event. Produces: who they are, what they care about, what to say, what to ask.
Input
Person/company + meeting context
Output
Focused briefing saved to knowledge/meeting-prep/
Detail
Runs WebSearch for recent activity. Saved file is also referenced by the calendar-intelligence daily task for its evening meeting check.
Trigger
"weekly review" / "Friday wrap"
Weekly wrap covering goal progress, LinkedIn content angles, stale context flags, and three prioritized focus areas for the coming week.
Input
Week's context files and automation logs
Output
knowledge/YYYY-MM-DD_weekly-review.md
Detail
Reviews progress against quarterly goals, surfaces the best content angles from the week's captured insights, and flags any context files that are becoming stale.
Trigger
"track this application" / "any follow-ups due"
Job search pipeline management using Linear as the authoritative source of truth, synchronized with the priorities context file.
Input
Application details + Linear state
Output
Updated Linear issues + context sync
Detail
Creates Linear issues per application, tracks status transitions, flags stale applications at 7+ days with no activity. Demonstrates using a project management tool as an AI-readable state machine.
Trigger
"/loop" / "start the loop"
Continuous intelligence heartbeat โ€” the core automation engine. Self-renewing, session-scoped, absorbs multiple other automation functions.
Input
All context and state files
Output
Updated command-center.html + console output
Detail
9-phase cycle, 2-hour interval via CronCreate. Registers its own cron job and renews it at 60h (before the 72h expiry). Absorbs: calendar prep, health nudges, weekly review elements. Cost: ~$0.17/cycle.
Trigger
"research X" / "deep dive on"
Iterative scored web research with quality threshold. Stops when findings are good enough. Flags gaps explicitly when not.
Input
Research topic + current priorities context
Output
knowledge/research/YYYY-MM-DD_slug.md
Detail
Scoring rubric: coverage (3pts) + relevance (4pts) + actionability (3pts). Score under 6: iterate with refined queries. Max 3 rounds. INSIST items surface immediately when a finding requires same-day action.
Trigger
"think ahead" / "what am I missing"
Anticipatory analysis across 5 domains. Surfaces non-obvious insights with specific actions โ€” not vague observations.
Input
All context + latest automation logs
Output
knowledge/thinking/YYYY-MM-DD.md + top 3 insights
Detail
5 domains: career trajectory, content opportunities, relationship intelligence, goal progress, system optimization. Every insight must include a specific action. Runs embedded in heartbeat Phase D or on-demand.
LinkedIn QR

Scan to connect

Vladimir Vorotilov