promptdojo_

The spec is the build — tickets an AI can execute without you in the room — step 6 of 8

Inside or outside the frontier

Back to the 758-consultant experiment, because it hands you a delegation rule and most people miss it.

The researchers didn't find that AI is good or bad at knowledge work. They found a frontier: inside it, big gains on speed, quality, and completion; outside it, a 19-percentage-point penalty on getting the right answer — inflicted mostly on people who trusted fluent output on a task the model couldn't actually do. The output looked equally confident on both sides. That's the trap. Fluency doesn't degrade at the frontier; correctness does.

You can't see the frontier, but you can probe for it. Before delegating a task to an AI builder, ask one question:

Does checkable ground truth exist for this task?

  • "Migrate these 40 report templates to the new format" — ground truth exists. The old templates are right there; the output either matches them field-for-field or it doesn't. A wrong answer is cheap to catch, so even if the task sits near the frontier, the failure mode is a caught failure. Delegate.
  • "Draft the integration plan for the client's undocumented legacy system" — no ground truth in reach. The facts live in a system the model has never seen and in the heads of two engineers who built it in 2014. The model will produce a beautiful plan anyway. Nothing in the output signals which parts are load-bearing guesses. Keep it human, or pair it: AI drafts structure, a human who can check reality owns every claim.
  • "Add retry logic to the export job" against a ticket with no acceptance checks — this one isn't a frontier problem, it's a spec problem. Delegating it wastes a build proving the ticket was vague. Spec-first: bounce it to the lint, fix the ticket, then triage again.

Three lanes, one router

That's the whole triage: delegate (ground truth exists, spec is tight), pair-review (no checkable ground truth — AI drafts, a named human verifies every claim before it counts as progress), spec-first (the ticket fails lint; nothing gets built until it passes).

Notice what the lanes protect. Delegate protects your throughput — inside the frontier the gains are real and large, and hoarding that work is how PMs become the bottleneck this chapter is named for. Pair-review protects your correctness — it's the 19-point penalty, pre-empted. Spec-first protects your week — every vague ticket you bounce at write-time is a rework loop you never enter.

The checkpoint at the end of this lesson makes you build the router — a few lines on top of the lint you already wrote. You'll use its logic on every ticket you write from now on, whether or not a script is running.