Same Batch of Models: Why Is Only Your Context Valuable?

Greg Isenberg posted a 49-item list on X, with a very blunt title: turning your company into the most AI-intensive company is the whole game. Original post: https://x.com/gregisenberg/status/2121331227009364340.

This list isn't like a prompt collection; it's more like breaking the company down into a queryable, versionable repository. The second brain is an Obsidian folder containing markdown files that record how you think, how you write, and how you make decisions. The company brain adds another layer: customers, calls, numbers, pricing, playbook, past decisions, and why those decisions were made at the time.

Greg's approach is clear. Use a voice tool like Wispr to feed in everyday thoughts, and have agents archive facts, stories, and decisions. Record calls and meetings with Granola, and send transcripts straight into the brain. Maintain a separate facts file that clearly states the true things agents are allowed to say, to reduce fabrication. Then maintain a voice file: 50 of your best posts, 20 posts you hate, and words you would never use.

Put these files in GitHub so every change has a history. At the top level, put a CLAUDE.md or AGENTS.md requiring agents to read the files before any task, not overwrite old content, and write clear dates. Cross-session memory must also be provided, otherwise agents start from zero every time. Consolidate data sources into one repository so ads, analytics, CRM, and payments can all be queried. Greg's exact point: models are the same for everyone, and only your context belongs to you. Accumulate context as much and as fast as possible.

The second half of the list is more execution-oriented. First install existing skills; Anthropic's skill library, Superpowers, Matt Pocock, and Addy Osmani are all worth checking out. Install marketing skills too. Rank every workflow in the company by whether it makes money or saves money. Do each process yourself first, narrate while doing it, record it, and then have an agent turn the recording into a skill: when to use it, steps, checkpoints, and mistakes to avoid. A skill should do only one thing; small and boring is more reliable than one giant prompt.

Use MCP to connect tools so agents can act, not just give advice. Gmail, Slack, Stripe, Linear, Notion, PostHog, and Supabase can all be connected. Agents also need a browser and their own computer to operate software without an API. Keep an orchestrator always online, one agent per project, and each agent runs sub-agents. Coding agents do the building, wrapped in a harness that makes them plan, test, and remember. Manage all agents in one place, with tasks, budgets, and logs. Skills can have triggers: run daily, run weekly, run when a new lead comes in, run when someone unsubscribes. Route models too: expensive ones for planning and checks, cheap ones for volume. Cost every task; if cost is higher than value, fix it or kill it.

The factory section is like a software pipeline. Spec agent, multiple build agents running in parallel via git worktree, test agent, review agent, and finally deployment. Design goes through the same line. Every annoying task becomes an internal tool that same day. Every customer request first becomes a pull request, which a human then reads. Review agents need explicit criteria, such as 'Would the best team in the world ship this?' If the answer is no, reject it. Keep 50 real cases and correct answers; when a new model is released, rerun them and switch to the winner. Humans keep approval, accountability for outcomes, and customer communication.

The loop section is even more critical. Every completed piece of work gets written back into the brain: what worked, what failed, what should stop. Support tickets, analytics, and social comments flow in every night. Every Friday, agents rewrite the playbook based on this week's results, and a human approves the diff. Every month, rank by profit per human hour, kill one thing and double down on another. Growth also runs through agents: research before spending a dollar, outbound automatically finds leads, enriches them, verifies them, and sends; the ad loop writes creative, runs it through the Facebook Ads API, kills losers, and scales winners. Every completed project becomes content: stories, numbers, prompts—anything. Every customer makes the brain smarter for the next customer.

Some replies are closer to real-world implementation than the list itself. One commenter asked who maintains all of this. That question is more realistic than item #1. Another said the 50 real cases are the real asset; models come and go, but standards remain. Another warned that every claim in the facts file should have a source, an owner, and an expiration date; otherwise an agent will faithfully repeat last quarter's pricing and say it with complete confidence. Another pointed out from an engineering perspective that LLM context windows are limited, and cramming everything into an agent is not optimal. Another joked that recording every call may most directly result in fewer invitations to meetings. Yet another said they had built a similar system and spent more time feeding it than selling.

What's truly hard to copy about this list isn't the tool names. Facts files go stale, playbook turns into a junk drawer of old decisions, recordings change the atmosphere of meetings, and agents confidently cite wrong information. Greg's proposed solution is versioning, writing dates, rewriting on Fridays, and killing tasks by cost. You can copy the 49-item structure; you can't copy the people who maintain the structure.