Autonomous Claude System

User sets goals. Claude orchestrates specialists. Quality over speed.

Updated: February 2026 | Architecture v3 (3-Prompt + Prompt Critic)

System Overview

👤 USER

Types rough prompt in Launcher UI or browser

▼

✨ PROMPT IMPROVE

Haiku polishes prompt automatically at queue time

▼

📋 PROMPT QUEUE

future_prompts table — GET /api/future-prompts

▼

🎯 ORCHESTRATOR

Reads queue, runs /build_prompt, manages worker lifecycle

▼

🚨 PROMPT CRITIC

Gates vague prompts

🔧 BUILD PROMPT

3-part: identity + task + tools

▼

🎨 web_designer

UI, dashboards

🔬 researcher

Deep research

🖥️ infrastructure

Deploy, SSH

🗃️ database

DB queries

📧 emails

Email processing

⏰ scheduled

Cron jobs

⚙️ fc_manager

Pipeline changes

▼

✅ REPLY & RECYCLE

Worker replies via send-prompt → orchestrator verifies → /clear → next task

Prompt Lifecycle

User submits rough prompt

Via Launcher UI (localhost:5111) or browser sync

launcher.py — Flask server

Prompt Improve NEW

Haiku automatically polishes prompt at queue insertion time. Stores original in prompt_original, improved in prompt_raw

improve_prompt_at_queue_time() in launcher.py

Prompt Queue

future_prompts table in Master_Claude.db. GET /api/future-prompts returns next prompt

Orchestrator reads & builds prompt

Runs /build_prompt to create 3-part prompt: identity (who the worker is) + task (what to do) + tools (MCPs, skills, learnings from DB)

build_3prompt_architecture() in launcher.py

Prompt Critic gates dispatch GATE

Claude Haiku semantic validation — blocks vague prompts from wasting worker sessions

prompt_critic.py

Worker launch

Specialist Claude launched in split pane (add_to_grid: true). Prompt includes CLAUDE.md, skills, domain context from DB

POST /launch → grid layout

Worker executes

Specialist works autonomously. Contextual hints inject just-in-time via hooks at tool-call time

Worker replies

Sends completion message via POST /send-prompt/{orchestrator_id}. Messages arrive as system-reminders on next tool call

pipe_inbox.py — inter-session messaging via named pipes

Orchestrator verifies & recycles

/verify → /goodbye (capture learnings) → /clear → send next task to same worker session

3-Prompt Architecture

Every worker receives a prompt assembled from three parts, pulled from Master_Claude.db:

Part	Contains
Identity	Who the worker is, constraints, design principles, session ID
Task	The actual work, reply-back instructions, messaging protocol
Tools	Skills, MCPs, learnings, domain context, available commands

# /build_prompt pulls from:
- specialist_identities  (who)
- specialist_contexts    (domain knowledge)
- skills table           (available skills)
- learnings table        (past experience)
- mcp_configs            (tool servers)
- hooks                  (contextual hints)

Specialist Workers

🎨

web_designer

UI, dashboards, HTML pages, Cloudflare Pages deploy

🔬

researcher

Deep research, web search, analysis

🖥️

infrastructure

Deploy, SSH, server work, tunnels

🗃️

database

DB queries, schema changes, migrations

📧

emails

Email processing (hansenemail)

⏰

scheduled_tasks

Cron jobs, recurring automation

⚙️

focused_claudes_manager

Launcher & pipeline changes

Workers launch as split panes in the grid. After /clear, any worker becomes any specialist — the prompt defines the role, not the project directory.

Key Components

Component	Purpose	Location
Launcher API	Flask server — sessions, grid layout, prompt queue, messaging	`localhost:5111`
/build_prompt	Assembles 3-part prompts from Master_Claude.db context	`launcher.py`
prompt_critic.py	Haiku-powered semantic validation — gates vague prompts	`claudequeue/`
pipe_inbox.py	Inter-session messaging via named pipes	`claudequeue/`
Contextual hints	Just-in-time reminders injected via hooks at tool-call time	`~/.claude/skills/contextual_hints/`
cls search / cls all	BM25 knowledge search across all FTS tables	`Master_Claude.db`
Grid layout	Orchestrator + workers in split panes (`add_to_grid: true`)	`launcher.py`
Master_Claude.db	Central knowledge — skills, learnings, errors, hooks, contexts	`D:/ClaudeQueue/`

API Endpoints

Session Management

`POST /launch`	Launch new Claude session
`GET /sessions/active`	List active sessions
`POST /send-prompt/{id}`	Send message to session
`POST /complete/{id}`	Mark session complete

Prompt Queue

`GET /api/future-prompts`	Get queued prompts
`POST /api/future-prompts`	Add prompt to queue
`DELETE /api/future-prompts/{id}`	Remove from queue

Worker Lifecycle

📨 Send Task

Prompt via /send-prompt

→

⚒️ Execute

Autonomous work

→

💬 Reply Back

Results to orchestrator

→

✅ /verify

Check work quality

→

🔄 /clear

Recycle for next task

Key principle: Launch once, reuse forever. Max 2 panes: orchestrator + 1 worker. After /goodbye captures learnings and /clear resets context, the same session handles the next task as any specialist.

Infrastructure

  ┌─────────────────┐     ┌──────────────────┐     ┌─────────────────────┐
  │  d_pc (Windows)  │     │  Cloudflare Pages │     │  londonserver (VPS) │
  │                  │     │                   │     │                     │
  │ launcher.py:5111 │───▶│  claudequeue.com  │     │  /api/* via tunnel  │
  │ Master_Claude.db │     │  (static files)   │───▶│  datasette, APIs    │
  │ Claude sessions  │     │                   │     │                     │
  └─────────────────┘     └──────────────────┘     └─────────────────────┘
        │                                                    │
        │  Wrangler deploy                    Cloudflare tunnel
        │  C:/Projects/ClaudeQueue/*.html     (only /api/*)
        ▼
  Static HTML → Cloudflare Pages → claudequeue.com

Architecture documented: February 2026 | Prompt lifecycle v3
← Back to Dashboard