Landmines in Context: How to Surface Buried Premises
The biggest risk to your project is a decision you didn't know was made. A buried premise can ruin your foundation - here's the rule you can give your AI collaborator to consistently and persistently surface risk.
Comes with a starter file for your Claude.
Henry Guyver · 8 June 2026
Effective AI collaboration is an orchestrated battle for truth. Excessive focus on hallucination reduction impedes progress, too loose a grip creates an output not worth retaining. You must hold truth like a greased egg - firm enough that it stays put, but not so firm as to have it slip your grip. Where do we actually have control as practitioners?
Your ability to execute with AI relies on two things: your ability to accurately and finitely describe what you wish to accomplish, in concert with your understanding of the context window - and how to manipulate it to shape the output. We'll save context window management for another time and focus on a failure within a single window - the buried premise.
When it comes to protecting accuracy with AI, the most dangerous thing isn't a bad turn, it's a missed assumption. Across a few turns a baseless assumption goes from a concept into galvanised 'fact', and has the ability to derail everything you're doing. If you aren't carefully monitoring the basis of the claims your AI is making, and validating that that basis is fresh, in terms of data recency, and accurate, in terms of your intended meaning, you risk poisoning the foundation of your project.
An example from my own work, and the impetus for the rule I now employ - Swimmable. Swimmable rates a bathing spot by how clean the water is. The obvious signal was distance - how close is the nearest sewage discharge. It was so obviously the right marker that the logic got silently baked into the project as truth. Except it's wrong. A discharge 5km upstream empties into the water you're swimming in while a closer one draining the other way never reaches you. Validate on distance alone and the map encourages swimmers into contaminated water. Every spot rating inherited the flaw.
It took enlisting an expert, Alex Lipp, an Earth Sciences lecturer at UCL, to spot the error. After he flagged it I asked Claude, 'what assumptions is our argument built on', and the premise was surfaced. We can't always have an expert on hand to spot these mistakes, so we need a system.
The challenge is that scouring your AI's thought process, reading all its internal deliberation is not sustainable long term and chaining "are you sure" prompts is tedious and has diminishing returns. What you need is a codified, persistent way to surface buried premises - the assumptions we are operating on turn after turn.
I've built that persistent ruleset for your AI collaborator, to help you protect your foundation and uncover the buried premise in your build. Give it to your Claude/Gemini/ChatGPT and it will internalise the rule and apply it thereafter.
Give this to your AI collaborator before your next build.