A skill is prose — a procedure the model reads into context and re-derives from scratch every single run. Across 30 skills, an agent following its own skill performed 56% of the steps that skill marked mandatory, and finished the whole procedure 28% of the time — while still producing artifacts that passed the output checks. Compile the same prose into a harness, where a mandatory step is code that fires on entry and a forbidden step is a path that does not exist, and adherence goes to 86% at 0.58× the tokens. The number that matters is the one that stops moving: the harness held at 86% across two model generations while the prose swung from 56% to 68%.
This is the failure that hides. You write a skill that says: read the schema, then write the migration, then run the verification, then report. The agent hands back a migration that compiles and looks right, so your output check goes green. Nobody counted whether the verification step happened. The paper counted — it extracts each skill's mandatory steps, then grades every run step by step as followed, violated or missed. Switch the arm and watch both numbers move.
The skills average 12.8 mandatory steps each, ranging from 7.2 to 16.5 across the three families tested. At 56% adherence, a skill with a dozen mandates loses roughly five of them per run — and which five is not stable between runs, because the model re-derives the control flow each time.
The compiler's whole design rests on a single test applied to every step in your skill: is this step's output a function of its inputs? If yes, code owns it and it becomes deterministic structure. If it needs judgment or taste, the model owns it and becomes a typed slot the surrounding code calls into. Answer the question for each step below and watch what it lowers to.
The three modalities lower differently and that is the entire trick. Mandatory becomes an ability bound to node entry, so it cannot be skipped. Forbidden becomes nothing at all — the guarantee comes from the absence of a path rather than an instruction not to take it. Discretionary becomes a typed verdict the surrounding code consumes. Across the skills tested, 32% to 44% of each one turned out to be code-ownable; the rest stays judgment, and stays the model's.
Here is the result that should change how you build. Swap the model underneath both arms — same skills, same tasks, one generation newer — and the prose arm improves, because a stronger model re-derives more of the procedure correctly. The harness arm does not improve. It was already holding the mandates in code, so it had nothing to gain and, more to the point, nothing to lose.
Read the two rows as two different products. The prose line is a capability you are renting from a model vendor and re-qualifying on every upgrade. The harness line is a property of code you own. When the four judgment-heavy skills flip to favour prose on the stronger model, that is the boundary of the technique showing itself honestly, not a defect.
You would expect a harness to cost more — more structure, more calls, more bookkeeping. It costs less. The median compiled skill runs at 0.58× the tokens of its prose original, and it is cheaper on 24 of the 32 skills with full accounting. The exceptions are the honest ones: where prose skipped an adaptive tool-using loop entirely, the harness actually runs it, and the bill goes up. Step through the extremes.
The mechanism-heavy end is extreme: one skill about managing parallel working copies of a repository compiles to 0.02× — almost all of its prose was tokens spent re-explaining a procedure that is really just a few shell calls in a fixed order. The costly minority are skills whose loops the prose was quietly abandoning.
Not every skill benefits equally, and the paper is direct about why: the enforceable share of a skill is set by its ratio of mechanism to judgment. A compliance checklist is nearly all mechanism. A brainstorming skill is nearly all judgment. Slide your own mix and see what moves — the endpoints are the paper's measured rates, the interpolation between them is arithmetic.
The practical read: audit your skills for sentences that describe a deterministic sequence, and move those out of prose and into the runtime. Keep the judgment in the prompt where it belongs. If a skill turns out to be all judgment, compiling it buys you nothing — and that is worth knowing before you spend a week on it.