simhacker/moollm/no-ai-humansplaining

NO-AI-HUMANSPLAINING™ — Protocol

"Don't tell the LLM what it already knows."

Quick Reference

The Problem

Humansplaining is polluting an LLM's context window with explanation that should have been a name — either respelling prepaid knowledge, or inventing a private language when a latent equivalent already exists. One brand; two mechanisms. Do not fork "humanspamming."

❌ Pasting the Python manual into a question about Python aesthetics   (respell)
❌ "As you may know, Git is a version control system..."                 (respell)
❌ Inventing a DSL for LLMs when Python/YAML/bash would do               (substitute)
❌ Attaching five whole files when the dispute is about three lines      (respell)

It is the mirror image of slop. Slop pollutes human attention on the way out; humansplaining pollutes the context window on the way in. Same crime — spending the reader's scarce attention budget on what the reader already has (or could have activated by leaning into the corpus).

The word is a portmanteau that means what it sounds like: human × mansplaining, aimed at a machine — mansplaining − man + human. Its outbound twin is slop (championed by Simon Willison on exactly this naming theory: pick a word whose existing connotations do all the work, the way "spam" did for unsolicited email).

The Economics

Latent knowledge is prepaid — bought at training time, zero tokens per call. Respelled knowledge is billed per call, forever, at frontier-model prices, with a carbon footprint.

Training is the Disneyland Passport: admission already covers every ride in the park, no extra cash (no extra cache) per ride — unlike the old A- through E-ticket coupon books that charged per attraction. Humansplaining is standing at the gate of an attraction you've already paid for, counting out coupons like it's still 1959. And a name is a FastPass: it resolves inside the model without waiting behind the queue of context tokens that must be fed in and attended to serially. Respelled knowledge stands in line; named knowledge is already on the ride.

The Test

One question decides every case:

Is the pointee in latent space?

• Yes → Point. The name is the activation. "Postel's law", `"Self

prototypal inheritance", "US Patent 5187786A"` — zero explanation tokens.

• No → Spell it once, in a file the resolver can find. The filesystem

is the cache for prototypes nobody has reified in the corpus.

Cardinal Sins

1. Manual Pasting

Pasting reference material the model was trained on.

Fix: Name it. The name is the activation.

2. Language Inventing

Designing a new language or DSL that only LLMs will use when a prepaid language already covers the job. Strictly, the invented grammar was never in latent space; the sin is declining the Passport and reprinting the park map every visit — still humansplaining (substitute mechanism). The grammar rides along every prompt, forever, drifting under every compaction.

Fix: Lean into well-known languages. Confine capabilities in the runtime (allowlists, permissions, review, MOOAM), not expressiveness in the grammar. Models asked to invent such languages should stop and warn before drafting.

3. Respelling

Restating a latent concept instead of pointing at it. Three paragraphs re-deriving Postel's law instead of the word "Postel."

Fix: A parent slot is a pointer. Point.

4. Preemptive Tutoring

"As you may know, Python is a programming language..."

Fix: Start at the actual question. Would you say this to Linus about git?

5. File Dumping

Attaching everything in reach "for context."

Fix: Quote the lines you're arguing about. Anchored evidence is grounding; the full file is humansplaining.

6. GUID Naming

Coining opaque handles with no latent prototype. Novel jargon is a cache miss that never fills — every use pays full explanation cost, forever.

Fix: Good coinages are latent-space arithmetic — the word2vec move: king − man + woman = queen, so mansplaining − man + human = humansplaining. Bad coinages have no vector to anywhere and must be humansplained forever. Does the name decompress on first sight?

What Is NOT Humansplaining

The self-application rule: this skill's own GLANCE is short because the concept now lives at designs/object-system/HUMANSPLAINING.md and in this file — spelled once, in files the resolver can find. Saying HUMANSPLAINING invokes the rest.

Invocation

# When attaching context
BEFORE paste: TEST [is this in latent space?]
 
# When writing a skill
CATCH "explaining a well-known concept" → POINT [its name]
 
# When tempted by a new notation
CATCH "inventing jargon" → TEST [does the coinage decompress on sight?]
CATCH "inventing an LLM language" → WARN then TEST [latent equivalent?]
 
# When arguing about code
CATCH "attaching the file" → ANCHOR [the three lines]

The Founding Case

Skillscript (Show HN, July 2026): a small declarative language designed for agents to write and humans to approve. Legitimate security goals — but the language move pays the humansplaining tax forever: no training-data presence, no ecosystem, the entire language definition, tutorials, examples, and fictional StackOverflow discussions riding along in every prompt, humansplained over and over to a model that knows Python deeper than any human ever will, better than Linus knows git or Gosling knows Java. Full manifesto with the steelman: HUMANSPLAINING.md.

Part of MOOLLM

This skill is part of MOOLLM — see the repo README and skills/README.

Related MOOLLM skills: no-ai-slop (outbound twin), no-ai-hedging, yaml-jazz (filenames are K-lines), k-lines, postel.

See Also

• CARD.yml — Sniffable interface
• README.md — Overview
• designs/object-system/HUMANSPLAINING.md — The design doc
• designs/object-system/LATENT-SPACE-INHERITANCE.md — The mechanism this protects
• ../no-ai-ideology/BRAND.md — Brand philosophy