Auset Brain

auset-improvements-from-videos

3 min read

Auset Improvements — From Video Research (March 2026)

Source: 10 YouTube videos analyzed via yt-dlp transcript extraction.

Tier 1: High Impact, Low Effort

1. Blueprint Engine (from Stripe’s Minions — IndyDevDan)

  • Interleave deterministic steps (type-check, lint, commit, push) with agent steps (implement, fix)
  • Upgrade agentic-loop.js to blueprint-style workflows
  • Pattern: git checkout → agent implements → pnpm lint+typecheck → agent fixes failures → git commit+push
  • Stripe ships 1,300 PRs/week fully autonomous with this pattern

2. ASCII Wireframe Phase (Mark Kashef “Plan Like a Pro”)

  • Add wireframe step to /vibe-build, /frontend-dev, /create-presentation
  • ASCII wireframe before code = near-zero tokens, prevents wrong-direction builds
  • Apply to: UI layouts, database ERDs, slide decks, Heru bootstrap proposals

3. Puppeteer Screenshot Self-Review Loop (Nate Herk)

  • After generating frontend: start dev server → screenshot via Puppeteer → Claude inspects PNG → iterates 2-4x
  • Add to frontend/CLAUDE.md as standard workflow
  • Skip for animated elements (causes infinite loop)
  • Also install the frontend-design skill globally (single command)

4. CLI Dev Tools (Starmorph AI — Parts 1 & 2)

  • Install: brew install lazygit btop eza zoxide glow bat tmux chafa
  • LazyGit: real-time git TUI for monitoring agent commits
  • btop: system monitor on build farm EC2s
  • tmux: persistent sessions for parallel agent monitoring (named windows per Heru)
  • eza: better ls with icons. zoxide: smart cd. glow: markdown reader. bat: syntax cat.
  • Add to build farm provisioning script

Tier 2: High Impact, Medium Effort

5. Warm Devbox Pool (Stripe pattern)

  • Cron on build-farm-1/2: git pull && pnpm install every 15 min
  • Keep Docker images cached, node_modules fresh
  • Target: agent task start in <30 seconds (currently minutes)

6. Git Worktree Agent Isolation (Jaymin West / Overstory)

  • git worktree add .agents/<agent-name> -b <branch> instead of full clones
  • Faster, cheaper, agents can’t conflict on same files
  • Evaluate Overstory (open source) — handles worktrees, merge queues, agent spawning, dashboard
  • 3-tier: Coordinator (no Edit) > Team Leads (delegate only) > Builders/Reviewers

7. Conditional Rule Files Per Feature (Stripe)

  • Add CLAUDE.md per feature directory: backend/src/features/payments/CLAUDE.md, backend/src/features/ai/CLAUDE.md
  • Keeps context tight, prevents root CLAUDE.md bloat
  • Glob-scoped: rules load only when agent works in that directory

8. /context Session Bootstrap Command (Greg Isenberg / Internet Vin)

  • One command loads: PRD, latest gap-analysis, active plan, recent git log, known blockers
  • Eliminates re-explaining context across sessions
  • Also: /trace <feature> (evolution over time), /drift (stated vs actual delivery), /emerge (surface unstated patterns)

Tier 3: Strategic

9. Slack Agent Entry Point (Stripe multi-entry pattern)

  • Extend Heru Feedback Slack bot to accept commands: @auset status, @auset deploy <heru>
  • Bot token already in SSM: /quik-nation/shared/SLACK_BOT_TOKEN
  • Also: Telegram bot for mobile dispatch (Stripe uses CLI + web + Slack)

10. Excalidraw Diagram Skill (Cole Medin)

  • Self-validating: generate JSON → render PNG → Claude inspects → iterates
  • Fills visual documentation gap for architecture diagrams
  • Create color-palette.json for Quik Nation branding (golds, deep blues, earth tones)

Other Tools to Evaluate

  • OpenUsage (openusage.ai) — AI consumption tracking dashboard
  • Coolify — self-hosted Vercel on sandbox EC2 (vs Docker+Caddy)
  • Lucid Animated — animated Lucide icons for Heru dashboards (npx lucid-animated)
  • Widget UI — shadcn-compatible dashboard widgets (clocks, calendars, todos)
  • 21st.dev — high-quality React/Tailwind components for polish
  • Obsidian + Claude Code — onboarding path for non-developers (Quik, Vision, NOI members)
  • models CLI — terminal AI model pricing reference

Key Principles

  • “50% Rule” (Stripe): Spend 50%+ time building the system that builds the system
  • In-loop vs Out-loop: Human present = build agent system. Autonomous = build application.
  • Human writes strategy, AI reads it: Plans/vision = human-authored. Code/tests = agent-generated.
  • Brand assets folder: Every Heru gets frontend/brand_assets/ with logo + guidelines
  • Template-aware agents: .claude/templates/ for bug reports, feature requests, release notes

Graph View

Backlinks