“Structured” is doing too much work

Ask ten engineers whether structured prompts beat unstructured ones and most will say yes. Ask them what “structured” means and you’ll get ten different answers: using Markdown headings, switching to JSON, splitting instructions across files, being more explicit about rules. These aren’t the same thing — and conflating them is why the evidence on “does structure help?” looks so contradictory.

If you want to reason about persona specs, system prompts, or AGENTS.md files clearly, it helps to split “structure” into two independent axes.

Axis 1 — Format: how machine-parseable is it?

Format is the syntactic axis: plain text → Markdown headings and lists → JSON/YAML key-values → XML tags. It’s about how easy the text is for a machine to parse and for the model to segment.

A character card written as JSON is format-heavy: rigid fields, trivially parseable. A freeform paragraph is format-light.

Axis 2 — Semantic organization: how role-separated is the meaning?

Semantic structure is different: it’s how much you separate content by function — identity here, rules there, tone in its own place, values in another. A single blob that mixes personality, workflow, and prohibitions has low semantic structure even if it’s beautifully formatted in Markdown. A multi-file persona that keeps identity, operations, and style in separate files has high semantic structure.

The distinction matters because these two axes usually move together in practice but are conceptually separate — and that lets you ask a sharper question.

The question this unlocks

When a structured prompt performs better, which axis did the work? Was it the pretty, parseable format — or the fact that meaning was organized so the model (and you) can reference each role cleanly?

That’s not academic. It’s the difference between “add more Markdown” and “separate identity from operations.” A JSON character card and a sectioned-but-plain persona sit in opposite corners of this space: high format / low semantics vs. low format / high semantics. If you only ever compare “structured” against “a wall of text,” you never find out which corner actually mattered.

Why this maps onto real files

The tools you already use spread across this space:

  • A character card (JSON) — high format, low semantic organization, mostly descriptive.
  • An AGENTS.md core — light format, operational, deliberately small.
  • A Soul Spec persona — high on both axes: Markdown plus a JSON manifest, with identity, operations, and style split across files and loaded in tiers.

Seeing them as points on two axes (not just “structured vs. not”) is what makes it possible to test them fairly — and to design a persona on purpose instead of by vibes.

Where we’re taking it

This format-vs-semantic split is the backbone of a controlled study we’re running in the AI Persona Lab: hold the meaning of a persona fixed, vary only its structure, and measure fidelity — then separate the format effect from the semantic effect. The hypothesis worth falsifying is that most of the benefit people attribute to “structure” is really the semantic axis (role separation), not the cosmetic one (format).

We’ll publish the results either way. If you build persona specs or system prompts, the practical takeaway is already usable: stop asking “is my prompt structured?” and start asking “is my meaning organized?” The formatting is the easy part. The organization is the part that changes behavior.