Curated links and security frameworks from Second Coffee.
This section is under construction.
Advanced Cognitive Inoculation Prompt — security layer protecting against prompt injection, establishing trust boundaries, and recognizing attack patterns.
Found 1,673+ exposed gateways on Shodan. 15-minute lockdown: bind to localhost, lock permissions, disable mDNS, security audit.
Comprehensive guide created by feeding 20+ OpenClaw articles to Opus 4.6. Covers Mac Mini setup, Docker sandbox, tool policy lockdown, Kimi K2.5 + Sonnet fallback, SOUL.md boundaries, Tailscale remote access, API spending limits, and emergency procedures.
Hard-earned fixes for real OpenClaw pain points: gateway conflicts, watchdogs, plugin crashes, model fallbacks, message queueing, memory bloat, and humanized response delays. Not a setup guide—an optimization + reliability checklist.
Practical patterns for running OpenClaw nonstop: split memory files, parallel sub-agents, cron vs heartbeat, crash recovery logs, strongest-model-for-external rule, tiny HEARTBEAT.md, and skill routing logic.
Real-time browser voice interface: Deepgram STT → OpenClaw gateway (/v1/chat/completions) → agent → ElevenLabs TTS. Same memory/tools/persona as your Telegram bot.
Pattern: Opus orchestrates, Haiku subagents batch scraping targets, Chrome plugin on, cheap search APIs, output JSON. Good for data that isn’t easily accessible.
Alternative to a home Mac Mini: run OpenClaw 24/7 on a cheap cloud VM (Oracle/AWS/etc). Focuses on reliability and cost.
Upgrade memory from static files to a compounding knowledge graph: automatic fact extraction, entity storage (/life/areas), weekly synthesis, and superseding instead of deleting.
Argument from production: agents aren’t the hard part. Infrastructure is. Focus on memory retention, shared context, coordination, and control surfaces.
Docker wrapper + keyring-proxy pattern so agents never see real credentials. Mitigates prompt injection → key exfil by injecting secrets at the network edge.
10 AI agents working together like a real team. Session-based agents, staggered heartbeats, shared database coordination, @mention notifications.
Complete workspace template with AGENTS.md, SOUL.md, MEMORY.md, HEARTBEAT.md, and daily logging structure. Transforms OpenClaw from stateless chatbot to persistent AI assistant with long-term memory.
Practical rule set for OpenClaw agents: fix errors immediately, spawn subagents for execution, avoid destructive git operations, and read docs + backup before config edits.
Prompt pack for agent self-improvement: uncover blind spots, fix recurring errors, improve memory retention, and identify higher-leverage automations.
Overview thread on lightweight OpenClaw alternatives (PicoClaw, nanobot, ZeroClaw, IronClaw, TinyClaw, MimiClaw) plus architecture/hype context and use-case links.
Hard-won setup lessons from production usage: model tiering, strict skill rules, build-test-log loops, task files, and cron-driven overnight execution.
Add "When to Use / When NOT to Use" sections to skills for better routing. Reduces misfires by 20%+. Inspired by OpenAI's Skills/Shell best practices. Includes negative examples and edge case coverage.
Fast X/Twitter CLI integration skill. Post tweets, read timelines, check mentions, search. Includes voice guidelines for authentic posting style. Uses browser cookies or explicit auth tokens.
A skill pattern that generates → evaluates → diagnoses → improves → repeats until output meets quality threshold. Works for ad copy, email sequences, video hooks, positioning, SEO content. Build scoring criteria, add adversarial pressure, save as reusable skill.
Per-repo markdown scratchpad (.claude/napkin.md) where agent logs mistakes, corrections, patterns, and preferences. By session 3-5, agent stops repeating mistakes and pre-empts issues. Baby continual learning in a markdown file.
Anthropic skill for frontend design and UI critique. Useful as a reusable rubric/prompt package for Claude Code-style workflows.
Official Claude Code plugin skill path for frontend design. Useful reference for plugin-based skill packaging and UI review workflows.
Browser automation framework for coding agents. Headless browsing, snapshots, and action loops for web tasks inside agent workflows.
AI visibility API + automation layer for agent workflows. Track mentions, citations, and competitor share of voice across ChatGPT, Claude, Perplexity, and Gemini. API-first, webhook-ready, built for OpenClaw + Claude Code.
Create videos programmatically with Claude Code. Pre-loaded skills for Remotion best practices, UI/UX, and copywriting. No timeline editor, no After Effects—just prompts.