Requirements Entropy, Not Token Count, Is the Real AI Coding Cost
Teams that skip requirement disambiguation will face agents generating plausible but wrong code, burning tokens and debugging cycles. The discipline of writing explicit, testable specifications becomes the competitive advantage in AI-driven development.
When an AI agent is given a twenty-page PRD and asked to break it into tasks, the result looks plausible but runs off the rails. The cause is not lack of capability but requirement entropy: every unstated assumption makes the model guess a probability distribution, and plausible guesses are worse than no answer. The article quantifies this with Shannon entropy and maps it to Cynefin's four domains, arguing that the goal of documentation is to lower the conditional entropy on the receiving side.
The core shift is from a reader-side to an author-side responsibility. Traditional PRDs rely on shared context among teammates; an AI has none. So every ambiguous phrase like "handle it per the old logic" or "exceptional cases are covered" must be chased down. The article proposes a decomposition pipeline: flag ambiguous spots, extract entities/states/rules/actions, query black boxes, align with existing code, and produce one task per testable rule. It also recommends using memory files (CLAUDE.md) as routing maps rather than encyclopedias, with explicit prohibitions on dependencies and keystone zones.
The article reframes an LLM's core value as de-conditioning: reducing the receiver's uncertainty rather than accelerating code generation. That inversion shifts the developer's skill set from syntax to specification.
The suggestion to use agent-based testing (performance/exception agents) as a way to encode tacit knowledge is a pragmatic move, but it assumes those agents themselves don't inherit the same tacit gaps, so human review on top remains necessary.
Placing the cost of entropy reduction on the writer side works for formalizable rules, but deceptive cases like chaotic incidents require constant human intervention; the claim that LLM can 'open black boxes' is optimistic until shown otherwise.