The Method
This is how the research is made.
The analysis comes from working inside the AI systems it describes. The method is what keeps that work traceable: portable context, receipts for the decisions that matter, and comparable setups across models. All of it lives in markdown files I control.
The method is open. Use it, adapt it, ignore it.
What the method is
Long-running work with AI breaks in a predictable way. Every new chat forgets. Every model starts from zero. Months later, it's hard to say where a claim came from.
The method addresses this with files, not software.
- The Thread. Memory and context carried together, from session to session and model to model.
- Core. A short file with the project's identity: what it is, who it's for, voice, rules, constraints. It travels with every conversation and rarely changes.
- Addendum. A running log of decisions, progress, and dead ends. It grows over time.
- Receipts. Addendum entries record where a consequential decision came from: who · why · source · date · model.
Threadbaire's own work has run this way since 2025. Articles, major decisions, and dropped claims are all logged.
How the research uses it
Predictions are filed before the event
A Tale of Two Earnings published six claims about Microsoft and Apple disclosures, each with pass/fail conditions, before the filings came out. The follow-up scored all six against the filings. All six held.
Claims get killed
Making claims testable only matters if failing claims come out. In Tools Multiply. Seats Collapse, a claim about declining Atlassian seats was dropped before publication. Its source was a competitor's marketing blog, and earnings data contradicted it. The claim would have strengthened the argument. The evidence didn't support it.
Comparable setups across models
For key questions, I give more than one model the same setup so their outputs can be compared. When they disagree, the disagreement is information.
Tools with defined jobs
Different AI tools do bounded work. One helps structure and draft. One argues the opposite side and tests the argument on a skeptical reader. One checks facts in a strict format: no narrative, sources cited, ambiguity flagged. The argument and the conclusions are mine.
The receipts are what make this visible. Without a record, you can't show what was checked, what was dropped, or what was said in advance.
Use it yourself
The templates are on GitHub. There's nothing to install, and they work with any model.
What's in the repository
README.md
How the method works, with tips.
Core_Template.md
Stable project information.
Addendum_Template.md
The decision log with receipts.
Extensions.md
Optional extras for specific kinds of work, such as dev logs, content metrics, and pipeline tracking.
How it works
Copy the templates
Rename them for your project.
Fill in the Core
Keep it short, because it travels with every conversation.
Log as you go
When you make a consequential decision or learn something durable, add an entry with a receipt.
Paste into new sessions
Start with the Core, then add Addendum entries when they're relevant.
A receipt looks like this
Decided to use Babylon.js for the 3D engine.
Receipt: Me + Claude · researching WebXR options · Babylon.js docs comparison · 2025-01-15 · Claude Sonnet
The fields are who · why · source · date · model. Use the ones that are useful.
A limit, from experience
On my active projects, plain files have carried roughly three to six months of work before outgrowing AI context windows. Your mileage will depend on how much you log. When it happens, split the log or move it into something you can query.
Where the method is going
Today the method is manual: files, copy and paste, receipts written by hand. That's deliberate. It proves the conventions work before anything is automated.
The next step is an automated version that works across frontier models and locally run open models, so the same record travels between them with nobody in the middle holding it.
Later, I plan to test the method in experiments with AI in spatial computing. The reason comes from the research: if value moves away from software and back toward the physical world, spatial computing is one of the first places where AI and the physical world meet.
Open resources
The research is published under CC BY-SA 4.0. Every article is available as raw markdown, with a Copy for AI button, an llms.txt file for AI discovery, and the full corpus on Hugging Face.
Threadbaire Server is an open-source database and API for when a log outgrows files. It's frozen at v1 as a reference implementation. RundownAPI is an open spec for an endpoint that teaches AI how to use an API from a single URL.
Commissioned analysis uses the same method. Every deliverable carries its evidence, its assumptions, and its failure conditions. Commissioned research →
Questions
Is the research written by AI?
AI is part of how it gets written, including drafting. The research, the argument, the claims and their verification are mine, and I'm responsible for every sentence that's published. AI tools do bounded jobs: structuring drafts, arguing the other side, checking facts. Factual claims such as figures, dates, quotes and filings are checked against primary sources. Interpretations and estimates are shown with their evidence and assumptions, so you can see what they rest on and disagree with them.
Do I need to install anything?
No. The method is plain markdown files. Copy them, edit them, and paste them into any AI conversation.
Does it work with ChatGPT, Claude, or other models?
Yes. Anything that accepts text works. That's the point of keeping it in files.
What's the difference between Core and Addendum?
Core holds what stays stable: identity, rules, constraints. Addendum is the running log of decisions and progress. Core changes rarely, while the Addendum grows.
How long do the files last?
In my experience with active projects, roughly three to six months before they outgrow AI context windows. Then you split the log or move it into something you can query.
Why markdown?
It's portable, readable, and works everywhere. Your context lives in files you control, not inside someone else's product. The Thesis spells out why that matters.
Is it free?
Yes. The templates, the server, and the RundownAPI spec are open source, and the research is CC BY-SA 4.0.