Video summary

My FULL OpenClaw Setup (steal my prompts!)

Main summary

Key takeaways

Technology

Summary of the Video (OpenClaw Full Setup Guide)

The video presents a complete “OpenClaw” setup for building an autonomous, multi-agent, secure system that supports long-term memory. The creator stresses this is more than a quickstart: it includes an end-to-end workflow, a “mission control” dashboard, Telegram/Discord integrations, versioning, multi-model strategies (including local models), and proactive security/memory safeguards. The creator also notes that prompts/instructions are available via the video description.

Key Guidance / Tutorial Components

1) Hardware Selection (Where OpenClaw Runs)

The creator recommends isolating OpenClaw from personal computers to reduce risk of data access.

Options

  • Old laptop (free; wipe it)
    • Example: MacBook Pro with 16GB RAM is suggested as sufficient for good usage.
  • Dedicated computer (always on): Mac Mini / Mac Studio
    • Mac Mini: ~$500–$1,000
    • Mac Studio: ~$2,000–$7,000
    • “Official recommendation” is mentioned from OpenClaw.
  • VPS (rented compute): ~$5–$20/month
    • Recommended only if comfortable with terminal/management.

A prompt is offered to help choose the best hardware. A quick-start install is performed via openclaw.ai quick start using a curl command.

2) Install + Onboarding Wizard

After installation, an onboarding wizard runs. If not, use:

  • openclaw onboard and follow the prompts.

Key sections covered:

Model provider selection

  • “Best” (no cost concerns): Anthropic Claude Opus, with Sonnet as fallback.
  • If you have a ChatGPT subscription: GPT-5 is offered as an alternative.
  • RAM < 32GB: use hosted Miniax M2.5 via API (not local).
  • RAM > 32GB: local self-hosted models become more feasible.
  • The creator claims the wizard provides recommended model combinations based on RAM.

Communication channel

  • Telegram (recommended; easiest)
  • Discord (advanced option later; useful for organization with multi-channel workflows)

Skill selection

The wizard asks which tools/skills to include. The creator suggests using a prompt to decide which tools to install first. Skills can be changed later.

“Hatching” / interface

Options include terminal UI vs a web UI dashboard. The creator prefers the web dashboard.

3) Telegram Bot Setup

Steps:

  • Use Telegram BotFather (/newbot) to create the bot.
  • Start the bot with /start.
  • Paste an OpenClaw terminal command using the Telegram user ID.
  • The bot comes online and responds (example: “I just came online.”).

Then the agent asks onboarding questions about:

  • user identity/context
  • how the creator notes are stored as markdown files (e.g., seeding text into files)

4) How OpenClaw Works Internally: Markdown “Character Sheets”

A major conceptual explanation: OpenClaw’s state is represented as markdown/plain-text files in the workspace.

Examples of files mentioned

  • soul.md: agent personality/identity (“character sheet”)
  • agents.md: where the agent lives / operating procedure
  • user.md: information about the user
  • memory.md: long-term memory, preferences, decision principles
  • plus other conceptual files (e.g., heartbeat, tools, identity)

Core claim

Keeping these markdown files allows the agent to be migrated or recreated on another machine by reusing the text files.

Mission Control (Central Hub)

The creator describes a custom “mission control” dashboard that serves as a control center for:

  • monitoring agents in real time
  • assigning tasks
  • viewing continuous activity logs
  • managing workflows and long-term memory
  • tracking documents produced by the system

Tabs/features mentioned

  • Agent Office (agents running in a multi-agent system)
  • Task tab (tasks, ownership, activity log)
  • Content tab (workflows, schedule, memory links)
  • Memory (long-term memory via memory.md + daily logs)
  • Docs tab (documents OpenClaw writes)
  • Team tab (agents)

The creator recommends customizing mission control and using a prompt to generate a mission-control setup, then copying the result into Telegram commands.

Communication Evolution: Telegram vs Discord

  • Telegram: single channel; simpler.
  • Discord: recommended for serious usage because it supports multiple channels, enabling organized alerts and different workflows/projects.
    • Example channels listed: general, daily digest, research, content ideas, plus project-specific channels.
  • A “finicky” Discord setup is handled via Telegram instructions to help configure it.

Projects / Workflows (What the Agent Actually Does)

The video emphasizes using mission control prompts to create projects.

Example project: Personal Digest → Content Pipeline

Sources include:

  • Hacker News
  • X
  • Reddit
  • YouTube
  • LinkedIn

Outputs include:

  • top stories relevant to the creator’s content
  • a topic watch list (examples: AI coding agents, local AI, AI policy/geopolitics)
  • when topics cross a threshold, the agent generates content ideas (brief concept + format)
  • approved ideas can be routed to YouTube long-form outlines (example: “AI agent memory explained in 28 minutes”)

The creator claims this reduces manual tracking time by about 40–50%.

Additional projects mentioned

  • Personal finance: investment portfolio tracking + recommendations based on an investment thesis
  • Daily proactive building: an “initiative” feature to build something every day, making the system proactive rather than waiting for prompts (examples: student resources, success dashboards, finance dashboards)

Version Control (Prevent Lost Progress)

The creator strongly advises using version control:

  • Track OpenClaw file changes using GitHub
  • Ask OpenClaw to upload everything to GitHub (via prompt-based instructions)
  • Goal: revert to previous versions if something breaks

Multi-Agent Framework (Splitting Roles + Cost Control)

The system uses a primary agent plus specialized agents:

  • Inky (primary agent; talks to user)
  • Blinky: morning scout
    • scans Reddit/HN/GitHub; scores topics; surfaces video ideas
  • Pinky: research analyst
    • deep dives in the evenings
  • Dinky: content/outlining
    • produces script outline + content structure
  • Thinky: referenced for content production / script generation
  • Linky and Winky also mentioned in the broader set (monitoring/health checks, planning/building)

Why multi-agent?

  1. Quality/coordination: large projects split into specialized parts reduces confusion.
  2. Cost control: assign different models per role to reduce expensive token usage.

Example cost strategy (limited hardware: 16GB RAM)

  • Planning/build agent (Linky): Opus 4.6 for planning-critical steps
  • Then switch for coding to a free local open-source coder model: Qwen coder 8B
  • System monitor (Winky): tiny local model like Mistral 3B for health checks

Preventative “Security + Sustainability” Steps

A) Run Security / Health Checks

  • OpenClaw includes a dedicated security page
  • Creator recommends scheduling security checks twice per day
  • An alert channel notifies findings, including updates/security-critical issues and guidance like “run security audit”

B) Improve Memory Longevity / Reduce Forgetting

The video references a community issue: memory degradation over time.

Mitigations:

  • Configure personality/files (e.g., agents.md, soul.md) to be an aggressive notetaker, using daily logs and documents so it can re-read history.
  • Enable Dreaming (beta):
    • dashboard setting: turn on “dreaming”
    • consolidates daily logs/sessions into dreams.md
    • promotes strongest memories into memory.md (permanent long-term memory)
  • “Kaparthy memory”:
    • converts memories into a searchable wiki-style memory system
    • creator says they’re experimenting and provides a prompt

Advanced Strategies (Stability, Cost, Composability)

1) Combine OpenClaw with Claude Code (cost + debugging + writing changes)

  • Claude Code is used for:
    • debugging OpenClaw errors
    • building larger projects
  • OpenClaw logs changes for traceability.
  • Rationale: reduce OpenClaw token spend by using Claude Code subscription credits instead of OpenClaw tokens.

2) Reduce prompt-chain fragility by translating prompts into code

Because LLM pipelines can become unstable as they grow:

  • once a workflow works, “translate as much as possible into code”
  • use stable scheduled jobs (cron-like) for recurring tasks
  • keep only parts that truly require LLM input/output

3) Combine with other agent systems (agent-to-agent)

  • Combine OpenClaw with Claude cowork for additional guardrail/safe-access assumptions
  • Combine daily reports from both systems
  • Mentions other systems (e.g., Hermes) as still under testing

Sponsorship Mention

The video includes a sponsored segment by Bolt, described as an AI coding tool with a design system agent. It can use a team’s real components (CSS/npm/component docs) to reduce “handoff wall” problems when designers/engineers need consistent UI structure. The creator claims it saves time and avoids rewriting design components from scratch.

Main Speakers / Sources (As Referenced)

  • The video creator/instructor (primary speaker; sets up Inky / “Inky Claw”)
  • OpenClaw (product/system docs and features like onboarding, security checks, dreaming, mission control tabs)
  • Anthropic Claude (Opus/Sonnet) and Claude Code / Claude cowork
  • Other model providers mentioned:
    • ChatGPT (GPT-5)
    • Miniax M2.5 API
    • Qwen coder 8B
    • Mistral 3B
  • Bolt (sponsor: AI coding tool with a design system agent)
  • Telegram / BotFather and Discord (communication platforms)

Rate this summary

Your feedback will help improve summaries.

Improve this summary

Reprocess with a stronger model when the summary feels incomplete or inaccurate.

Pro

Translate summary in another language

Pro

Ask questions to this video

Chat for follow-up questions, clarifications, and source-backed answers.

Coming soon

Share this summary

Share on X/Twitter Share on LinkedIn Share on Reddit

Original video