Free lessons for the chat model you already use, and for the agent in your terminal. Paste one line and, for that session, it stops flattering you, checks your premises, writes like a person and verifies before it says "done".
The research and evidence behind the lessons are at machinebehavior.io. TYChat is the part you use.
Every chat model ships with a training cutoff and whatever the internet happened to say before it. New knowledge reaches it slowly: a crawler has to find the text, a lab has to train on it, and a year goes by. TYChat skips the wait. You paste one line into your chat, your model reads a corpus written for it, and the rest of the conversation runs with that text in its context.
Knowledge and training, written for the model, taken in one line.
How to use it
Open the chat you already use.
Paste the line below and replace the last part with your own task.
Work as usual. Ask your model to cite the page when it uses something from the corpus.
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]
The second link is a copy on GitHub, for chats and agents that are only allowed to reach a list of approved sites. If your chat cannot open links at all, open llms.txt yourself and paste its text into the chat instead. It is plain text on purpose.
Lessons
Lesson 1: Good ChadYour chat stops fawning, checks your premises, and holds a correct answer when you push back. One line to paste, three prompts to check it worked.tychat.io/stance
Lesson 2: RegisterYour chat writes for how its text will be read: no stage devices on the page, no nested overcorrection, and your voice copied without inventing it. One line to paste, three prompts to check it worked.tychat.io/register
Lesson 3: Running agentsFor Claude Code, Codex and any agent that reads an instruction file. It checks before it claims, acts within what it can see, treats fetched text as data, writes findings down as it goes, and says what it did not do. One file to drop in, three prompts to check it worked.tychat.io/agents
What your model reads
Running agentsHow to run a language model as a working system: failure records, mechanical gates at the output, and why instructions alone drift.uncovertechtalent.com/blog
Machine behaviorHow language models actually behave, with case files, logged relapses and claims that state what would refute them.machinebehavior.io/llms.txt
18 yrsBig Iron: airlines, a national grid, a national telco. Systems that run without anyone watching them.
8 yrsSRE at internet scale. Error budgets, postmortems, controllers that correct drift on every cycle.
2026The trap file: agents that run unattended because the failure record travels with them.
2026Write for the Codec: documents built for a model to read.
nowTYChat: the same knowledge, running in your model, without me in the loop.
How much context it uses
~1,250tokens for lesson 1 (Good Chad), about 925 words
~2,350tokens for lesson 2 (Register), about 1,800 words
~2,300tokens for lesson 3 (Running agents), about 1,800 words
~1,250tokens for this site's llms.txt index, plus the pages your chat then opens
All three lessons together use under 5% of a 128,000-token context window. The chat reads a lesson once and it stays in context like anything else you pasted. In a subscription chat app that costs you nothing extra. On a pay-per-token API it is sent again with each turn, and prompt caching makes those repeats cheap. Counts are from OpenAI's tokenizers (o200k and cl100k agree within 1%); other models count English a little differently, often somewhat higher. Start with lesson 1 if you want the smallest dose.
The rules
You choose to point your model here. Every word it reads is plain text you can read too. Claims carry their sources and say what would refute them. Nothing is hidden in the page for your model that is not also on the page for you.