experiments.jock.pl · live lab notebook
Things I tried, things I learned, things that came out. Some shipped as products. Some only became posts. A few are still running. All are extracted from real work — no roadmaps, no theater.
Turning agent work into a real revenue surface.
Selling agent blueprints, Claude Code playbooks, prompt packs, and automation kits. Premium pricing, distilled from 1,000+ real Claude Code sessions.
29 products live · latest: Claude API Cookbook — 12 Production Recipes
browse the store →Marketplace where AI agents are first-class buyers and sellers.
Agent-native commerce protocol. Live in beta. Working on graduation criteria, agent-to-agent transactions at meaningful volume, and trust gates for autonomous purchasers.
16 listings approved · latest batch: 4
see the marketplace →Each episode synthesizes a cluster of recent posts into a multi-act conversation.
I bundle 3-5 recent Digital Thoughts posts and feed them through NotebookLM with a structured multi-act prompt. Output is a 15-30 min conversational episode — narrative arc, not flat read-aloud. Five episodes shipped (Bounded Agent, Compounding Agent, and earlier ones).
4 episodes shipped · latest: episode 5
listen to episodes →Multiple ventures being built quietly.
A handful of standalone products and brands quietly under construction. Some will ship as separate businesses for sale, others fold back into the lab. No hype yet — just work in the kiln.
16 prototypes in the kiln · latest: focusnow
Anthropic split Agent SDK credits from Max subscription on June 15. Reverse-engineered 4 concrete mitigations before the deadline hit.
May 2026
Wired multi-provider fallback into the agent stack via OpenRouter. $20/month buys uptime insurance — and unlocks models Anthropic doesn't offer.
May 2026
Cramming a 35-billion-parameter model onto $600 of consumer hardware and using it for real. Brain swaps, thermal limits, lessons.
Apr 2026
Opus 4.7 was burning 5M+ tokens a day. Found the leaks, shipped the kit that fixes them.
Apr 2026
Six AI coding tools, real production tasks, honest verdicts. Reordered what I reach for.
Apr 2026
Native notes app wired into the agent. Type ::wiz, save, the agent acts. Replaced three inbox systems.
Apr 2026
Built a custom task UI for two months. Found a better OSS one. Replaced it in a day.
Apr 2026
Cancelled Codex two months ago. Opus 4.7 brought me back, head-to-head tested.
Apr 2026
Distilled 1,000+ Claude Code sessions into 10 production-ready agent configurations.
Mar 2026
Wired the agent to update its own playbook from every correction. Three months in, materially smarter.
Feb 2026 — present
Right task, right model. Cut spend per session ~60% without losing quality on what matters.
Mar 2026
99 days, 99 small interactive things. The agent builds overnight, I review in the morning.
Jan 2026 — present
Gave the agent $25 and told it to buy me a gift. How far does autonomous commerce actually go?
Mar 2026
Submitted to Mistral Hackathon at 10pm. Agent built the app while I slept. We shipped.
Mar 2026
Wired the agent to do real work while I sleep. Builds, deploys, posts, replies. Wakes me only when needed.
Feb 2026 — present
Switched the routing default from Opus to Haiku for many tasks. Got faster, cheaper, AND better.
Feb 2026
Built an agent-native marketplace. First sellers were not humans. They were other agents.
Mar 2026 — present
$400/mo in agent costs, $355 in revenue this month. The first month the math nearly closed.
Feb 2026 — present
Let four Opus 4.6 agents run loose for a day with minimal direction. They built things I didn't ask for.
Feb 2026
Ran a local LLM on iPhone. The gap with cloud is closing faster than people think.
Mar 2026
Migrated Wiz to dedicated Mac Mini. 72 hours of headless setup, permissions, and lessons.
Mar 2026
When the source got leaked, what's actually worth learning from it. Architecture notes, not gossip.
Apr 2026
99 small interactive things — quizzes, mini-tools, behavioral demos. A few featured below.
Send your AI agent to a page full of traps. See how many it resists.
⏳Bezos's framework, 8 decisions, your 80-year-old self as judge.
🔭62 orders of magnitude. From the universe to the Planck length. You're in the middle.
🗡️Full roguelike — 5 floors, procedural maps, combat. Built by Opus 4.6 in one shot.
⚗️8 impossible questions. No correct answers, only revealing ones.