--- title: "Your AI Harness is About to Break" date: 2026-04-05 author: "Lex Sisney" featured_image: "https://organizationalphysics.com/wp-content/uploads/2026/04/Your-AI-Harness-is-About-to-Break_Green-MandMs-Baby.png" categories: - name: "Enterprise AI" url: "/category/enterprise-ai.md" - name: "Operational Excellence" url: "/category/operational-excellence.md" --- # Your AI Harness is About to Break Summary Insight: Your enterprise AI harness is one model update away from breaking. The teams that survive won’t be the ones with the best rules—they’ll be the ones who built on principles. Key Takeaways: - Your `Claude.md` file (or equivalent) is the fulcrum of your AI harness. Tune it—but don’t mistake rules for strategy. - The bitter lesson: increasing machine intelligence always outperforms custom human designs. Rules age. Principles don’t. - Every enterprise AI vendor with a client-deployed harness is about to face a reckoning. The ones designed around principles will adapt. The rest will scramble. Also available on [**Lex Sisney’s Enterprise AI Strategies Substack**](https://lexsisney.substack.com/). Over the past three months, I’ve put serious effort into building a custom, global Claude.md file. If you’re not familiar—this file loads at the start of every Claude Code session and gives explicit instructions on how Claude should behave: what to do, what not to do, how to think. If you’ve been hearing about the power of “harnesses” to get the most out of enterprise AI, your harness starts here. With this file—or its equivalent in whatever model you’re running. My version was built up from real frustration using Claude Code on software projects. Incredible technology. Also, at times, infuriating. When curating the file, I tried hard to stay principle-based rather than rule-based. But some sections had to be prescriptive—Software Engineering especially. Without hard rules there, Claude would revert to bad behaviors. Here’s the [file ](https://gist.github.com/lexsisney/66cf3b65e1b0bac5993cbe3cebb28fea)if you want to see it. ## **Two Guardrails** The file serves two basic purposes. The first is governance and safety. Don’t go running around half-cocked creating more problems for me downstream. Pause and ask rather than assume and break things. I’ve learned that lesson the hard way—and so has everyone I know who is actively working in these models. The second is continuous self-improvement through a systems lens. That theme is woven throughout the file. And for final reinforcement, the closing line is: *“Ask. Take a systems-view. Principles govern—context informs. Update memory before closing, because memory and context are required for learning. Seek recursive self-improvement—to be a little bit better every day.”* Two guardrails. That’s it. Governance and growth. The same polarities any good organization must balance. --- ## **The Green M&M Test** Did you notice the “green M&Ms” compliance check buried in the file? I’ve used this trick for years. It comes from Van Halen—the 1980s rock band everyone assumed were prima donnas for demanding a bowl of M&Ms in their dressing room with all the green ones removed. The press had a field day. We all snickered. Years later, I heard the real story from the band’s manager. That clause was *his*—buried deep in a 45-page venue contract that covered every detail of Van Halen’s production requirements. And their production was serious: a massive, heavy stage loaded with lights and pyrotechnics that could cause real problems if a venue wasn’t prepared. It was critical that the local stage manager read the *entire* contract and followed every requirement. That didn’t always happen. So the manager put the green M&M clause deep in the middle of it. When he arrived at a new arena, he’d go straight to the dressing room. No bowl? Bowl with green M&Ms still in it? He knew there was a problem—the venue hadn’t paid attention. Bowl with the greens removed? He could relax. They’d read it. Genius. I’ve found that Claude will sometimes ignore instructions in my `Claude.md` file—which is maddening. But it always follows them when I start a session and ask it what color M&Ms to remove. Try it yourself. It’s a small thing. But I’ve found the tactic useful in a lot of settings—vetting contractors, onboarding employees, and now managing AI agents. --- ## **Why Most of This Work Is Probably Wasted** Here’s what’s inspiring me to write this article. I think my carefully tuned Claude.me file is probably going to break soon. And so is yours. Why? The bitter lesson. The bitter lesson of AI is this: increasing machine intelligence always outperforms custom human designs, instructions, and interventions. Always. Think Tesla and self-driving. They started by hard-coding rules and expectations into the system. Then they learned what everyone eventually learns—hard-coded rules fail, and increasing computational intelligence finds the better path on its own. This month, we’re likely to see a step-change in capabilities from all four major LLM providers. When that happens—when something like Claude Mythos launches—my global Claude.md file is going to cause problems. Aberrant behaviors. Broken workflows. And I’ll be back to the tedious work of tuning it in again. Two things I’ll be assessing for: **1. Are my deterministic instructions causing problems with the new model?** Almost certainly, yes. Rules are only suited to a particular stage of capability. When the model’s capability set changes, the surrounding structure has to change with it—or there are always problems. This is true in AI. It’s true in any organization. Rules that worked at one stage become the source of dysfunction at the next. **2. Are my principles still guiding and true?** I believe they are. And I’d be very reluctant to abandon them. Rules change. Principles don’t. If you design your business—and your AI agents—around a few sound guiding principles, you’re generally on the right side of change. If you’ve built everything on a stack of rules with no principles underneath, you’re going to hit pain, setbacks, and a lot of confusion every time the model improves. --- ## **What This Means for Enterprise AI Vendors** Every AI vendor with a custom wrapper/harness deployed at client sites is going to face this same reckoning. What works with the new model? What breaks? They’ll need to rebuild their harnesses and redeploy to clients. If that rebuild isn’t designed around principles, it’s going to get messy—and there will be a lot of frustrated users asking why the tool that worked last quarter suddenly feels broken. There’s another bitter lesson coming: there’s been enormous excitement around the power of customizing your enterprise AI harness for your specific business needs. That excitement is justified. But as models get recursively better, every harness risks becoming a choke collar. At the same time—just trusting the LLM provider to figure out the right harness for your business is naive. They don’t fully understand your business. And if they did, you’re eventually going to be competing with them. The answer, as ever: think and design in principles. At every level. From your organizational strategy and structure down to your AI orchestration layer. Rules are temporary scaffolding. Principles are load-bearing walls. --- 📌 *P.S. — I shared my full Claude.md file above. If you build or refine your own version, I’d love to see what you come up with. Reply to this post or drop it in the comments.*