The Labour Doesn't Disappear — It Hides

Gabriel Gonzalez's "A sufficiently detailed spec is code" makes one argument that actually holds up, and one test that doesn't.

The argument: if you make a specification precise enough to reliably generate working code, you've essentially written code. He earns this by going into OpenAI's Symphony project and pulling out what the SPEC.md file actually contains — database schemas in prose, backoff formulas written as pseudocode, and a section the authors titled "Cheat Sheet (intentionally redundant so a coding agent can implement the config layer quickly)." That last one is worth pausing on. A specification document that includes a redundancy layer explicitly for the model is not a specification. It's a codebase wearing a suit to its own job interview.

That observation does something most AI commentary won't: it opens the file and reads it, rather than arguing about the category.

The part that doesn't hold up is the Haskell test. He gives Claude Code the Symphony spec, asks it to generate a Haskell implementation, it fails, and this is offered as evidence that spec-to-code generation is fundamentally broken. But Haskell is a low-resource language that's underrepresented in model training data — the author even acknowledges this in a footnote before waving it away. A poorly controlled experiment can still point at a real problem. This one does. But it can't carry the weight it's being asked to carry.

The section worth reading carefully is the one on slop — specifically the claim that the industry conditions driving agentic coding adoption are the same conditions that make careful specification writing impossible. If you're treating the spec as a productivity shortcut, you'll produce a document that looks like a specification and functions like a rushed first draft. The tool doesn't create that failure mode. It just gives organisations a new way to call it a methodology.

The agentic coding hype will pass. The underlying pressure to move faster than your thinking allows won't.