📺 Channel: VelvetShark

[Why your OpenClaw agent forgets everything (and how to fix it)](https://www.youtube.com/watch?v=oN__gKJnPls)

Channel: VelvetShark

Summary:

• Here's a summary of the video "Why your OpenClaw agent forgets everything (and how to fix it)":

Key Takeaways

• The primary issue discussed is that OpenClaw agents frequently forget instructions due to context compaction, which is a mechanism to manage the agent's limited memory window.
• A comprehensive memory architecture is essential for agent reliability, encompassing multiple layers and strategic configurations.
• Understanding the "4 memory layers" and "3 failure modes" is crucial for diagnosing why an agent forgets instructions.
• Specific techniques, such as a "pre-compaction memory flush" and a "/compact timing trick," can significantly extend the lifespan of agent instructions.
• The way data is structured into files and the retrieval methods used (e.g., hybrid search vs. QMD) are fundamental to memory persistence through compaction.
• Effective memory management not only improves agent performance but also leads to cost savings by reducing redundant API calls.

Main Arguments

• The central thesis is that the common problem of OpenClaw agents losing instructions is directly tied to how the system handles context compaction. The video argues that this isn't an unsolvable flaw but a consequence of not fully leveraging or understanding the existing memory architecture.
• The guide presents a multi-faceted solution, breaking down memory into distinct layers and processes (pre-compaction flush, manual discipline, file architecture, retrieval) that must be managed individually and in conjunction.
• By implementing specific configurations and practices outlined in the video, users can build agents that retain instructions for longer periods, leading to more consistent and reliable performance.
• The video emphasizes that adopting a deliberate "memory protocol" and understanding advanced retrieval mechanisms like QMD are key to overcoming the limitations of standard context compaction.

Notable Quotes (Interpreted from direct statements and emphasis)

• "OpenClaw's #1 problem: your agent forgets its instructions after context compaction."
• "This is the guide I wish existed when I started." (Indicating the guide is a comprehensive, foundational resource).
• "The pre-compaction memory flush most users never discover." (Highlighting a powerful, underutilized feature).
• "The /compact timing trick that gives new instructions maximum lifespan." (Suggesting a practical, impactful optimization).
• "Compaction vs. pruning (completely different)." (Emphasizing the distinction between two related but separate memory management concepts).

Important Nuances

• Distinction between Compaction and Pruning: The video clarifies that context compaction (managing the active context window size) and session pruning (removing older or less relevant memories) are fundamentally different processes, though both relate to memory management.
• Pre-compaction Memory Flush: This is presented as a critical, often overlooked step that occurs before context compaction, allowing specific information to be preserved.
• `/compact` Timing Trick: The timing of the `/compact` command is shown to be a tunable parameter that can be exploited to grant newly provided instructions a longer duration within the agent's active memory.
• File Architecture for Survival: The organization of files is not merely for tidiness but is designed to "survive compaction by design," implying specific naming conventions or directory structures are recommended.
• Retrieval Strategies (Track A, A+, QMD): The video discusses different methods for retrieving information from the agent's knowledge base, ranging from basic tracking to more advanced techniques like QMD (Quantum Memory Dispersal), each with its own trade-offs.
• API Cost Savings: A key benefit of a well-architected memory system is the reduction in API calls, as the agent can recall information rather than needing to re-query or re-infer it, thus saving money.
• Memory Protocol: A structured approach to memory management is proposed, with a recommendation to add this protocol to the `AGENTS.md` file for consistent application across agents.
• Hybrid Search vs. QMD: The video contrasts OpenClaw's built-in hybrid search capabilities with the more advanced QMD for searching the entire knowledge base, suggesting QMD for deeper, more complex retrieval needs.

Published: 2026-03-06T21:36:42+00:00

[50 days with OpenClaw: The hype, the reality & what actually broke](https://www.youtube.com/watch?v=NZ1mKAWJPr4)

Channel: VelvetShark

Summary:

• Here is a summary of the video "50 days with OpenClaw: The hype, the reality & what actually broke":

Key Takeaways

• OpenClaw, after 50 days of consistent daily use, proves to be a versatile self-hosted AI agent with a broad spectrum of practical applications across personal and professional life.
• The long-term reality of using such an agent involves a balance between significant benefits and inherent challenges, including system stability issues, operational costs, and the ongoing need for human oversight.
• Security is a paramount concern for self-hosted AI agents, requiring proactive and diligent mitigation strategies.
• Effective integration with existing tools and robust community support are crucial for maximizing OpenClaw's utility and overcoming hurdles.

Main Arguments

• The true capabilities, limitations, and practical value of a self-hosted AI agent like OpenClaw are best understood through extended, consistent real-world usage over an extended period (e.g., 50 days).
• While AI agents offer advanced automation and problem-solving, they are not yet fully autonomous and require ongoing management, troubleshooting, and occasional human intervention (e.g., memory compaction, task babysitting).
• The real power of OpenClaw is unlocked through its integration into daily workflows, infrastructure management, and knowledge management systems, rather than as a standalone tool.
• Users must critically assess the cost-benefit analysis and clearly define their specific use cases ("what do I use it for?") to derive meaningful value from their AI agent setup.
• Proactive and robust security measures are indispensable when deploying and managing self-hosted AI agents.

Notable Quotes/Phrases

• "This is my 50-day OpenClaw review after running a self-hosted AI agent every single day."
• "This is a real 50-day field report."
• "20 real OpenClaw use cases from daily life"
• "Discord channel architecture + per-channel model routing"
• "Markdown-first workflows with Obsidian + semantic search"
• "What actually breaks (memory, compaction, browser automation)"
• "Security risks and how I mitigate them"
• "The "what do I use it for?" problem"
• "Tasks that still need babysitting"
• "The verdict: should you use OpenClaw?"

Important Nuances

Extensive Use Cases Demonstrated (20 total)

• Daily Automations: Morning Twitter briefings for day organization, daily AI art generation for e-ink displays, and automated self-maintenance tasks like updates and backups performed overnight.
• Always-On Checks: Background health checks that successfully caught critical issues, such as a forgotten Netflix payment failure.
• Research & Content Generation: Use cases include a research agent that can spawn multiple parallel sub-agents, querying YouTube analytics in plain English, and a rapid URL summarization skill (`/summarize`).
• Infrastructure & DevOps: Practical applications in server migration, managing zombie processes, and even enabling "coding from phone" capabilities (though deemed unsuitable for production environments).
• Daily Life Integration: Streamlining email triage (via a draft-only mode), managing shared calendars with a spouse through WhatsApp, transcribing voice notes into actionable outputs, and handling various personal reminders (e.g., coffee shops, weather, product reminders).
• Discord, Knowledge Management & Creative Workflows: Migration from Telegram to Discord was highlighted as a significant upgrade, with advanced Discord bookmarking replacing dedicated tools like Raindrop. Integration with Obsidian for semantic search across over 3,000 notes is a key workflow. Other creative uses include a WordPress honeypot, generating Excalidraw diagrams, and ongoing development for Home Automation with Home Assistant.

Technical Limitations & Breakage Points

• Significant challenges were encountered with "memory loss and context compaction," impacting agent reliability over extended periods.
• Browser automation was identified as a common area where functionalities tend to break.
• Certain complex or critical tasks still necessitate human oversight and "babysitting" to ensure successful completion.

Operational Realities

• The "cost reality" of running self-hosted agents is substantial, underscoring the need for cost optimization and careful consideration of the investment.
• The common user hurdle of the "what do I use it for?" problem is addressed, emphasizing the importance of defining clear, valuable applications for the agent.

Security Considerations

• The video explicitly discusses "security risks" associated with self-hosted agents and details practical mitigation strategies employed by the speaker.
• The existence of specialized resources like the "ClawHub security check skill" is noted.

Workflow Enhancements & Community Resources

• Detailed architectural insights into Discord setup, including per-channel model routing.
• Emphasis on "Markdown-first workflows" facilitated by Obsidian and semantic search.
• The active community, showcased on Clawdiverse.com, and recommended "starter pack" workflows are highlighted as valuable resources for new users.

Published: 2026-02-20T22:34:46+00:00

[I cut my OpenClaw API bill by 80% with one config change](https://www.youtube.com/watch?v=fkT41ooKBuY)

Channel: VelvetShark

Summary:

• Here is a summary of the video "I cut my OpenClaw API bill by 80% with one config change":

Key Takeaways

• Many OpenClaw users are unknowingly overspending on API costs due to default configurations.
• Implementing "multi-model routing" or "model tiering" is presented as a solution to reduce API expenses significantly (50-80%).
• This strategy involves directing different types of tasks to specific AI models based on their cost and performance characteristics, ensuring quality for critical tasks while using cheaper models for less demanding ones.
• The video provides guidance on configuring multi-model routing for heartbeats, sub-agents, and main tasks within OpenClaw.
• A ready-to-use JSON5 configuration file is offered for easy implementation.
• The `/model` command is highlighted as a tool for real-time cost control.
• A free cost calculator is available to help users estimate their potential savings.
• The discussion includes reasons why relying solely on free tiers might not be the optimal approach.

Main Arguments

• The Problem: The default setup for OpenClaw often utilizes more expensive AI models for all operations, regardless of task criticality, leading to inflated API bills. This is exacerbated by the vast cost differences between various available models.
• The Solution: Model tiering or multi-model routing is proposed as the fix. This involves creating a system where requests are intelligently routed to different models. For example, simple, repetitive, or non-critical tasks (like heartbeats or basic queries) are assigned to highly cost-effective models, while complex, critical, or high-priority tasks are directed to more powerful (and typically more expensive) models. This ensures that users pay only for the capabilities they truly need for each specific task.

Notable Quotes

• "If you're running OpenClaw, there's a good chance you're burning money right now without realizing it."
• "cut your API costs by 50-80% - without losing quality on the tasks that actually matter."
• "That's a 60x difference between the cheapest and most expensive option." (This quote refers to the significant price disparity between AI models).

Important Nuances

• Model Pricing Variability: The video mentions a significant price difference (e.g., a "60x difference") between AI models. While specific pricing examples were provided for models like Gemini 2.5 Flash-Lite, DeepSeek V3.2, GPT-5, and Claude Opus 4.5, some of the quoted prices appeared to be malformed or unusually low (e.g., `0.00` for Claude Opus 4.5). These should be treated as illustrative of the cost spectrum rather than exact figures and may require verification from current sources.
• Configuration Details: A specific JSON5 configuration file is offered, intended to be easily copy-pasted, detailing how to implement the multi-model routing rules.
• Runtime Control: The `/model` command is presented as a valuable utility for users to actively manage and control model usage and costs on the fly.
• Strategic Use of Tiers: The video suggests that while free tiers exist, they might not always be suitable for all OpenClaw operational needs, implying that a tiered approach combining paid and free models strategically is more effective.
• Scope of Optimization: The configuration and routing strategies discussed are applicable to distinct components within OpenClaw, specifically heartbeats, sub-agents, and main processing tasks.

Published: 2026-02-04T00:21:47+00:00

[OpenClaw (Clawdbot) use cases: 9 automations + 4 wild builds that actually work](https://www.youtube.com/watch?v=52kOmSQGt_E)

Channel: VelvetShark

Summary:

• Here's a summary of the video "OpenClaw (Clawdbot) use cases: 9 automations + 4 wild builds that actually work":

Key Takeaways

• OpenClaw (Clawdbot) is presented as a practical AI agent framework that can be integrated into daily life for tangible results.
• The video showcases 9 real-world automations currently in production and 4 more advanced "wild builds" that demonstrate the system's broader capabilities.
• A core message emphasizes that AI agents are ready for daily use and that commitment to their integration unlocks productivity and creative potential.
• Security is a critical component, with dedicated guardrails discussed to ensure safe and reliable operation.
• The platform is designed to be useful for a wide range of users, including business professionals, developers, and hobbyists/creators.

Main Arguments

• AI agents have evolved beyond theoretical concepts into practical tools that real people are actively using in production environments today.
• Committing to running an AI agent daily can lead to significant improvements in workflow efficiency and creative output.
• OpenClaw supports a diverse spectrum of applications, from routine tasks like email management and reporting to complex development workflows and hardware integrations.
• Robust security measures are fundamental to building trust and enabling the widespread adoption of AI agents in everyday life.

Notable Quotes

• "what's possible when you commit to running an AI agent in your daily life"
• "9 automations that actually work, 4 wild builds that show the ceiling, and the security guardrails that keep it all safe."
• "Every example here is something real people are running in production right now."
• "If you already know what OpenClaw/Clawdbot is, this video shows you what to actually DO with it."

Important Nuances

Target Audiences

• Business users: Applications include email triage, customer support, and investor communications.
• Developers: Use cases cover PR reviews, multi-agent workflows, and general code assistance.
• Hobbyists/creators: Examples span smart home automation, IoT projects, and creative hardware.

Content Breakdown

• 9 Automations: Morning briefing, Email triage, Homelab daily report, Slack customer support, PR review → Telegram, Multi-agent dream team, Camera trigger automation, TRMNL "Moment Before", Daily AI app builder.
• 4 Wild Builds: These are presented as more advanced or experimental examples that highlight the system's potential and "ceiling." Specific details are not provided in the transcript snippet.
• Key Features Discussed: The video touches upon "Skills & memory," "Security guardrails," and practical considerations such as "Cost & getting started."
• Community Engagement: The video encourages users to share their own "cool builds" by adding them to "Clawdiverse" or commenting, indicating a community-driven aspect.
• Practical Focus: The content is geared towards demonstrating actionable uses of OpenClaw, assuming the viewer already has a basic understanding of what the platform is.

Published: 2026-01-31T02:46:01+00:00

[Compound Engineering: the AI coding workflow that actually learns](https://www.youtube.com/watch?v=4hLJ62m3OqI)

Channel: VelvetShark

Summary:

• Here's a summary of the video "Compound Engineering: the AI coding workflow that actually learns":

Key Takeaways

• Most developers use AI for coding to achieve immediate speed gains, overlooking the greater potential for long-term improvement.
• "Compound Engineering" is a workflow that treats every AI interaction as an investment, leading to a system that becomes progressively faster and smarter over time.
• This approach contrasts with standard AI sessions by ensuring that knowledge gained from code reviews and bug fixes is retained and compounded, rather than being lost at context limits.

Main Arguments

• The current approach to AI in coding is suboptimal because it focuses only on immediate productivity, failing to build a sustainable learning system.
• Compound Engineering advocates for a workflow where every development step—from brainstorming to code review—contributes to the AI's continuous learning and knowledge accumulation.
• This compounding knowledge base allows the AI system to improve its performance and efficiency for all future sessions, creating an exponential benefit.

Notable Quotes

• "Most developers use AI to code faster today. But they're missing the bigger opportunity - building a system that makes them faster tomorrow, and every day after."
• "Compound Engineering is a workflow where every piece of work is an investment."
• "Every code review teaches the system something new. Every bug fix becomes a rule it won't break again."
• "this knowledge persists and compounds."

Important Nuances

• The video emphasizes building an AI coding system that learns and evolves, rather than treating AI as a mere tool for individual tasks.
• The demonstrated workflow (brainstorm → plan → work → review → compound) is crucial for facilitating this continuous learning process.
• This methodology directly addresses the transient nature of typical AI coding sessions by establishing a persistent, ever-growing knowledge base.
• A specific "Compound Engineering plugin" is available to facilitate this workflow.

Published: 2026-01-23T22:37:55+00:00

[Claude Code for data analysis: 500,000 rows without writing code](https://www.youtube.com/watch?v=Ttf_IHlpZJk)

Channel: VelvetShark

Summary:

• In this tutorial, I walk through a complete data analysis workflow using Claude Code. No theory, just real data, real prompts, and real results you could send to your business today.
• Article with all the prompts to copy-paste: https://velvetshark.com/data-analysis-with-claude-code
• 📊 What we build:
• Excel report with KPIs, top products, top customers, and country breakdown
• Professional charts ready for presentations
• Customer cohort analysis with retention heat map
• Reusable script that regenerates everything with one command
• 🔑 What I cover:
• Setting up Python environment with Claude Code (one-time, 2 minutes)
• Preventing AI hallucination with verification steps
• Keeping "messy" data (returns, cancellations) for business insights
• Writing effective prompts for data analysis
• Spot-checking results to ensure accuracy
• 📁 Dataset used: Online Retail Dataset from UCI Machine Learning Repository
• https://archive.ics.uci.edu/dataset/352/online+retail
• 🛠️ Tools: Claude Code (https://claude.ai/code)
• You don't need to know Python. You need to know what questions to ask your data and how to describe what you want. If you're good with Excel, you already have those skills.
• ---
• 0:00 What we're building
• 1:07 The dataset: 500K rows of real retail data
• 2:56 You don't need to know Python
• 3:39 One-time environment setup
• 4:58 Loading data and preventing hallucination
• 5:52 Handling returns and cancellations
• 7:27 Building the reusable report script
• 10:05 The generated Excel report and charts
• 13:06 AI-generated business insights
• 14:06 Data verification (critical step)
• 16:43 Cohort analysis and retention heat map
• 17:31 Recap
• 19:31 Bonus: Jupyter Notebook in 2 minutes

Published: 2026-01-16T17:50:52+00:00

[ClawdBot (OpenClaw): The self-hosted AI that Siri should have been (Full setup)](https://www.youtube.com/watch?v=SaWSPZoPX34)

Channel: VelvetShark

Summary:

• Siri still can't remember what you told it yesterday. But what if you could have a personal AI assistant that actually remembers your conversations, works inside Telegram or WhatsApp, and can message YOU first with daily briefings?
• Meet ClawdBot - a self-hosted AI assistant created by Peter Steinberger that runs 24/7 on a cheap VPS and connects to the messaging apps you already use.
• In this video, I show you step-by-step how to set up ClawdBot on a Hetzner VPS (~$5/month), including:
• • Creating and configuring your server
• • Installing all prerequisites
• • Running through the onboarding wizard
• • Connecting Telegram as your interface
• • Adding skills to extend capabilities
• • Real use cases: proactive briefings, research summaries, and more
• ⚠️ IMPORTANT: This is a real agent with real powers. Read the security docs before connecting sensitive accounts.
• 🔗 LINKS:
• • Aricle with all commands you can copy-paste: https://velvetshark.com/clawdbot-the-self-hosted-ai-that-siri-should-have-been
• • ClawdBot Docs: https://docs.clawd.bot/start/getting-started
• • GitHub: https://github.com/clawdbot/clawdbot
• • ClawdHub Skills: https://clawdhub.com/skills
• • Security Guide: https://docs.clawd.bot/gateway/security
• • Discord Community: https://discord.com/invite/clawd
• • Hetzner: https://www.hetzner.com/cloud
• 💰 COSTS:
• • VPS: ~$5/month (Hetzner)
• • AI Model: Claude Pro $20/month OR API usage-based
• 0:00 - What Siri should have been
• 0:38 - What is ClawdBot?
• 3:19 - Why you need a $5 server
• 5:04 - Setting up your VPS
• 11:01 - Installing ClawdBot
• 17:37 - Your first message
• 21:51 - Use cases & what's possible

Published: 2026-01-09T19:38:41+00:00

[Stop prompting Claude Code - let it interview you (the "spec" workflow)](https://www.youtube.com/watch?v=ob9WWuYlS5Q)

Channel: VelvetShark

Summary:

• Stop feeding Claude Code vague prompts. Let it interview YOU instead.
• This technique went viral when Anthropic engineer Thariq shared it. In this video, I break down exactly how it works and show you the full workflow from a 1-sentence spec to a working CLI tool.
• 📋 THE PROMPT (copy & paste):
• ---
• Read @SPEC.md and interview me in detail using the AskUserQuestionTool about literally anything: technical implementation, UI & UX, concerns, tradeoffs, etc. but make sure the questions are not obvious. Be very in-depth and continue interviewing me until it's complete, then write the spec to the file.
• ---
• 🔧 THE WORKFLOW:
• 1. Start with a minimal spec (even one sentence)
• 2. Claude interviews you with probing questions
• 3. After 10-40+ questions, you have a detailed spec YOU control
• 4. Start a fresh session and implement the spec
• 📁 SLASH COMMAND (/interview):
• Create this file at ~/.claude/commands/interview.md
• ---
• allowed-tools: AskUserQuestion, Read, Glob, Grep, Write, Edit
• argument-hint: [spec-file]
• description: Interview me to expand the spec
• ---
• Here's the current spec:
• @$ARGUMENTS
• Interview me in detail using the AskUserQuestion tool about literally anything: technical implementation, UI & UX, concerns, tradeoffs, etc. but make sure the questions are not obvious.
• Be very in-depth and continue interviewing me until it's complete, then write the spec back to $ARGUMENTS.
• ---
• 🔗 Original tweet: https://x.com/trq212/status/2005315275026260309
• 🔗 Copy-pasteable code: https://velvetshark.com/stop-prompting-claude-code-let-it-interview-you
• 💬 Try this on your next feature and drop a comment: How many questions did Claude ask you?
• 0:00 How it all started
• 0:36 The problem with prompting AI
• 1:14 The technique: AI interviews YOU
• 1:46 The 4-step workflow
• 2:14 Demo: One sentence to full spec
• 3:06 32 questions later...
• 4:13 Implementing the spec
• 4:44 Does it actually work?
• 5:21 Why this works so well
• 6:24 Pro tip: Add anti-goals
• 6:47 When to skip this technique
• 7:09 Creating a /interview slash command
• 8:37 Try it yourself

Published: 2026-01-02T14:57:31+00:00

← back to home