The Prompt FilesWhite papers, 2017 → now
Every landmark paper from Attention Is All You Need to this month's releases, in order, with one honest sentence on why each one matters. Filter by topic or search.
Browse the timeline
AI/ML research · prompt engineering · since 2023
The Prompt Files
A working record of how large language models came to be steered by text: the white papers that mattered, the techniques that survived, and the tools to put them to work.
—
curated papers and reports
—
years of continuous trail
7
eras, from attention to agents
0
trackers, logins, or paywalls
Start here
Three ways in, depending on what you need today.
Prompt library
Guides and copy-ready templates: fundamentals, reasoning models, context engineering, structured output, agents, and low-resource languages.
Open the libraryPrompt Constructor
Fill in role, goal, context, constraints, and format. Get a structured prompt that follows the current best practice for the model you are using. Built in your browser, with an optional rewrite by Claude.
Construct a promptLatest additions to the trail
The most recent entries on the white papers page.
What this site was meant to be
The Prompt Files started in 2023 as a research notebook on LLMs and GPTs, with a mission to make prompting rigorous and to push language models toward communities that the big training sets leave out. The first version shipped with three papers and a lot of placeholders.
This version fills them in. The paper list is now a continuous record, the library has real guides instead of titles, and the constructor works without a server. The mission is unchanged: cutting-edge research, explained plainly, with a bias toward what you can actually use.
What changed since 2023
- 2023GPT-4, open weights (LLaMA), tool use, jailbreaks and prompt injection. Prompting is a set of tricks.
- 2024Million-token context, open models catch up, the first reasoning model (o1), prompt caching, MCP. Prompting gets a survey and a taxonomy.
- 2025Open reasoning models (DeepSeek-R1), RL on verifiable rewards, computer-use agents, alignment-faking and unfaithful-reasoning results. "Context engineering" replaces "prompt engineering".
- 2026Long-running agents, evaluation and oversight as the research frontier. See the 2026 section for what is current.