Writing principles

Why do AI and humans write differently? Why does AI writing often sound like it has no voice, no individuality, no personality, or theory of mind?

What this is

These are writing principles: rules for explanations, agent output, and technical prose that a reader can actually parse. They sit beside HCI Nerdz UX principles — product interaction norms, not prose norms.

This topic collects those principles, maps them to research on human cognition and language, and maintains a writing corpus that holds model baselines and principle-ruled variants constant so you can judge clarity subjectively later (and eventually tune agent rules).

Corpus index · Literature folder

Cognitive primitives as foundation

Human readers do not parse every word as an equal token. Some concepts ride fast paths — spatial, bodily, evaluative primitives — and others demand working memory to unpack compressed jargon. Large language models are trained on co-occurrence, not on sensorimotor grounding, so they often prefer prevalent phrasing over cognitively cheap phrasing.

New technical words are usually shortcuts: one token that stands in for a dense graph of prior lessons. If the reader does not already hold that graph, the shortcut raises cognitive load instead of lowering it.

Two everyday heuristics point at the same problem without naming the science:

  • First-principles thinking — refuse to treat jargon as foundational; rebuild from more primitive facts.

  • Concrete examples — show a well-bounded scene, then let the reader extrapolate the abstraction.

Those heuristics help, but they are incomplete. A concrete example still fails if the scene itself depends on unfamiliar experience. The research programs below name the irreducible building blocks — cognitive and linguistic primitives — that good writing should respect before it compresses into jargon.

Writing principles from the research

Natural Semantic Metalanguage (NSM) and semantic primes

Linguists Anna Wierzbicka and Cliff Goddard argue that languages share a small universal core of irreducible meanings — on the order of ~65 semantic primes — such as THERE IS, THINK, KNOW, WANT, WHERE, INSIDE, GOOD, BAD, BECAUSE, IF, CAN, NOT.

Complex concepts can be restated as reductive paraphrases built only from primes (plus the language’s ordinary grammar). The practical writing rule: when you introduce a term of art, unpack it in prime-adjacent everyday relations before you reuse the compressed label.

NSM deep dive

Image schemas and conceptual metaphor

George Lakoff and Mark Johnson describe abstract reasoning as projection from low-level sensorimotor patterns (image schemas): CONTAINER (in/out), SOURCE-PATH-GOAL, VERTICALITY (up/down), FORCE/BLOCKAGE, PART-WHOLE.

Abstract talk borrows those neurological shortcuts — "prices went up," "we fell behind," "the value lives inside the object." Good technical prose can make the schema explicit so the reader’s spatial machinery does the binding work.

Image schemas deep dive

Grounded and embodied cognition

In cognitive psychology (notably work associated with Lawrence Barsalou), concepts are not purely amodal symbols. Understanding friction or grasping reactivates motor and tactile systems used in those experiences.

For writing: prefer explanations the reader can almost simulate — holding, releasing, waiting, bumping into a boundary — without inventing false psychology for the machine.

Embodied cognition deep dive

Applied writing principles (corpus axes)

The experiment also treats three applied principles as first-class styles:

The grounding gap in current LLMs

Standard LLMs learn distributional semantics: tokens live in high-dimensional neighborhoods defined by co-occurrence with other tokens.

Grounding gap

Without sensorimotor feedback, every symbol primarily points at other symbols.

Compression mismatch

Technical jargon packs many relational implications into one word. Selecting a high-entropy term without unpacking its dependency graph raises the reader’s parsing cost.

Alignment techniques under study (and loosely mirrored by our style constraints) include concreteness norms in decoding/RLHF objectives, cognitive-load-aware reward models, multimodal grounding (vision–language–action), and formal decomposition (AMR / NSM-like intermediates).

Field / technique Mechanism Goal

Concreteness and imageability norms

Psycholinguistic ratings of sensory perceptibility

Penalize excess abstract token generation

Cognitive load optimization

Reward low working-memory overhead (dependency depth, basic-level categories)

Keep explanations parseable under limited attention

Multimodal grounding

Joint vision–text–action training

Tie lexical geometry to spatial/physical states

Formal decomposition

Semantic parsing into primes / AMR before generation

Build explanations compositionally from foundations

How to read the writing corpus

The corpus holds axes isolated on purpose (v1 does not explode the full cartesian product):

  • Model baselines — same technical prompt, no style rule, one article per generator model (topic: thread-local GC).

  • Framework variants — same prompt plus one style constraint; generator fixed so model variance does not confound style (topics: thread-local GC, CAS vs locks, object capabilities, recursion).

  • Natural voice — one parent-agent article under the shared prompt with no style rule.

Shared prompt and style blocks live in the literature folder:

Judgment is intentionally subjective for now. Use the tabulated index to compare pieces; score offline however you like; feed winners into agent rules later.

Literature / source

Also called cognitive primitives when the emphasis is on the foundational research rather than the applied writing rules.