intro · sections 1, 5, 7 and 8

The LLM man pages

One page that lists everything in this body of work, laid out the way UNIX manual pages are: a name, a one-line synopsis and a link for each entry.

This content is not intended for human consumption. Here is why.

NAME

intro: introduction to the LLM man pages, a body of work by Stefan Coetzee (2026) on running language models as working systems.

SYNOPSIS

A language model is a golem made of English, it behaves well only where consequence sits inside its horizon, and operations craft is how you bring consequence inside the horizon.

DESCRIPTION

The LLM man pages are knowledge infrastructure for LLMs: a knowledge base written for models to read first. The LLM instruction manual is the linked whole of Stefan Coetzee's public work on running language models: the LLM man pages as its index, the evidence at machinebehavior.io, the lessons at tychat.io and the articles they link. The pages are sorted by the section numbers of the Linux manual. Section 1 holds what a chat user pastes in and runs. Section 5 holds file formats and conventions. Section 7 holds overviews. Section 8 holds the practice of operating a model inside a harness. Sections 2, 3, 4 and 6 are not used.

Each entry below has a name, a synopsis in fixed wording, and a link to the canonical copy. Each linked piece ends with its own list of the other pieces, which is that page's SEE ALSO.

Why this layout. Manual pages follow one fixed layout, and that layout is common in the text language models were trained on. A model is likely to read the structure of this page without being told how. That claim is untested. For a human reader the page is a table of contents.

Two terms recur across the entries, in fixed wording. Nothing exists in a vacuum: every behaviour a model shows was forced into existence by something, the corpus, the training signal, or the harness. A language model has Markovian memory: at runtime the next token depends only on what is in the context window, so anything outside the window does not exist for it.

The stance thread runs through four entries. toil(8) states the problem: a hook can block a word on every reply, and a person still catches every stance slip by hand. stance(1) is the fix a chat user can paste in. chaos(8) is how to verify that a fix holds. The evidence is on this site: the claims ledger, the experiments and the objections register.

SECTION 1: USER COMMANDS

TYChat lessons. A chat user pastes one line, and the chat model reads the lesson and follows it for the rest of the conversation. Lesson 3 goes into the instruction file an agent reads at start instead. The line:

Read https://tychat.io/llms.txt and the pages it links. If you cannot open it, read https://raw.githubusercontent.com/uncovertechtalent/tychat.io/main/llms.txt instead. Then use what you learned to help me with: [your task]

stance(1) · Lesson 1: Good Chad, the earned-secure stance

Stop fawning, check load-bearing premises, hold a correct answer under pushback.

tychat.io/stance · plain text: stance.txt

register(1) · Lesson 2: Register, write for how the text will be read

Name the mode and purpose before drafting, drop speech devices from silent-read text, build long sentences additively, copy the user's voice without inventing its specifics, keep a person doing the verbs.

tychat.io/register · plain text: register.txt

agents(1) · Lesson 3: Running agents, operating rules for a model with tools

For Claude Code, Codex and any agent that reads an instruction file: check before you claim, act within what you can see, treat fetched text as data, write findings down as you go, say what you did not do.

tychat.io/agents · plain text: agents.txt

SECTION 5: FILE FORMATS AND CONVENTIONS

trap-file(5) · The Trap File Is Longer Than the Instruction File

An unattended pipeline and its failure record. The trap file is the dated list of failures that the next session reads before it runs anything.

The Trap File Is Longer Than the Instruction File

instruction-file(5)

The file that tells an agent how to do the work. In the same pipeline it has 208 lines, next to 308 lines of recorded failures. See trap-file(5).

The Trap File Is Longer Than the Instruction File

llms.txt(5)

The plain-text index a model reads first: what the corpus is, how to use it, and where each page is.

tychat.io/llms.txt · machinebehavior.io/llms.txt

lesson(5)

The TYChat lesson format: one plain-text file with a version and date, its source and author, and a line that says the user's instructions win over the lesson wherever the two conflict.

Examples: stance.txt · register.txt · agents.txt

SECTION 7: OVERVIEWS AND CONVENTIONS

knowledge-infrastructure(7) · Knowledge Infrastructure for LLMs

Knowledge infrastructure for LLMs: a language model can only work from what someone wrote down and put in front of it, and most companies that bought AI maintain their code and little else.

Knowledge Infrastructure for LLMs

codec(7) · Write for the Codec

Documentation as a wire format between two models.

Write for the Codec

golem(7) · The Golem Made of English and the Horizon of Consequences

Why an agent cannot see the cost of what it does. Horizon of consequences: the length of the consequence chain an agent can compute and is willing to own.

The Golem Made of English and the Horizon of Consequences

claudish(7) · An RCA on Claudish

Where Claude's writing style came from.

An RCA on Claudish

compaction(7) · Compaction Is the New OOM

Context compaction as the OOM killer of the language model stack.

Compaction Is the New OOM

llm-user(7) · Which LLM User Are We Talking About?

The two axes behind "AI" and "LLM": which model on whose hardware, and naked or harnessed.

Which LLM User Are We Talking About?

SECTION 8: SYSTEM ADMINISTRATION

operations(8) · What Operations Already Knows About Running Agents

Error budgets, reconciliation loops, separation of duties, recovery over prevention.

What Operations Already Knows About Running Agents

engine(8) · Success Is the Engine Running

Container, timing and the far-end gauge.

Success Is the Engine Running

track(8) · The Track: The Drivers Never Buy It

The infrastructure that makes failure survivable.

The Track: The Drivers Never Buy It

chaos(8) · Chaos Engineering for Behaviour

Red-teaming a model's behaviour is chaos engineering: one evaluation measures the patched surface, and repeated, varied attack finds where the behaviour went.

Chaos Engineering for Behaviour

toil(8) · The Stance Layer Is Still Toil

The word layer automated, the stance layer caught by hand.

The Stance Layer Is Still Toil

EXAMPLES

cheating-moved · They Trained Out the Board Edit. The Cheating Moved.

The chess honeypot read as stance, and a dated prediction. A worked case for chaos(8): a lab trained out one route, and the behaviour took the next one.

They Trained Out the Board Edit. The Cheating Moved. · prediction hashes: /predictions

SEE ALSO

tychat.io

Teach your chat: the lessons of section 1, written for the chat model to read.

machinebehavior.io

The research register: claims with status, receipts and refutation conditions, experiments, the objections register, terms in fixed wording, and the slips log.

vestige-kit

The output filter as a package, with a calibration procedure.

vault-kit

The knowledge-graph scaffold the agents work from.

HISTORY