Fabric

What it is

Fabric is an open-source framework for augmenting humans using AI, made by Daniel Miessler. It keeps AI prompts as Patterns, organized by the job they do. The README lists jobs such as extracting the most interesting parts of a YouTube video or podcast, summarizing an academic paper and explaining code. Others rate a piece of content to see if you want to read or watch all of it, or turn bad documentation into usable documentation. You run a Pattern from the command line, call it through an HTTP server, or copy its text into any AI app. This listing is Mindket’s setup notes for the fabric program, its server and its web app.

1. You send text, a web address or a YouTube link to fabric
2. fabric adds a Pattern, a Markdown prompt for one job
3. It sends both to the model provider you chose
4. The answer prints, streams or saves to a file

Who it suits

People who work from a terminal and repeat the same text jobs, such as summarizing articles, checking claims, pulling the key points from a video or explaining code. Developers who want those prompts behind an HTTP address can run the built-in server, and anyone who only wants the prompts can read them on GitHub with nothing installed. Skip it if you want no setup. Fabric needs a model, from a provider you have an account with or from a local model such as Ollama, or Apple Foundation Models on macOS 27 or later.

Where to use it

Fabric is a program for macOS, Linux and Windows. It talks to the model provider you set up, so the same Patterns can run against different models.

• Native providers in the README: OpenAI, Anthropic (Claude), Google Gemini, Ollama for local models, Azure OpenAI, Amazon Bedrock, Vertex AI, LM Studio and Perplexity
• OpenAI Codex, which the README describes as ChatGPT/Codex subscription OAuth via private backend, and Claude Code, which uses your Claude subscription through the local claude CLI
• Many OpenAI-compatible providers, among them Groq, OpenRouter, Together, DeepSeek, Mistral and LiteLLM
• An HTTP server, with an Ollama-compatible mode that shows each Pattern as a model
• A Svelte web app and a Streamlit app
• Any other AI app, by pasting a Pattern’s text into it

How to set it up

Get it saves a zip of Mindket’s setup notes in your downloads folder, with a page for each platform and feature from the maker’s README and docs. Start with the one-line installer, then run the setup command.

1. Run the installer for your system. It puts the fabric program in ~/.local/bin, or in %USERPROFILE%\.local\bin on Windows.
2. If that folder is not on your PATH, the installer prints how to add it.
3. Run the setup command. It sets up your directories and keys and asks which provider you want.
4. Start with fabric --help or fabric --listpatterns, as the installer page suggests.

curl -fsSL https://raw.githubusercontent.com/danielmiessler/fabric/main/scripts/installer/install.sh | bash
fabric --setup

On Windows, run the PowerShell line below, then fabric --setup. The README also gives winget install danielmiessler.Fabric.

iwr -useb https://raw.githubusercontent.com/danielmiessler/fabric/main/scripts/installer/install.ps1 | iex

Homebrew, the Arch Linux package and Scoop name the package fabric-ai. With Homebrew or the Arch package, the README says to add the alias line to your shell startup files so the fabric command works.

brew install fabric-ai
alias fabric='fabric-ai'

How to use it

Pipe text in and name a Pattern. The README’s examples read the clipboard with pbpaste on macOS. For Windows it says to use Get-Clipboard, and for Linux xclip -selection clipboard -o.

pbpaste | fabric --pattern summarize
pbpaste | fabric --stream --pattern analyze_claims

The first line summarizes the clipboard, and the second runs analyze_claims and streams the answer as it arrives. The list command shows every Pattern. The help text says --readpattern prints one to the terminal.

fabric --listpatterns
echo "test input" | fabric --dry-run -p summarize

The dry run shows what would be sent to the model without making an API call. Add -o to save the answer to a file and -c to copy it.

Example

You want the key ideas from a long video without watching it. Install yt-dlp, which Fabric uses to fetch the transcript. The YouTube page gives the Homebrew line for macOS and says to use your system’s package manager elsewhere. Then give fabric the video’s address and the summarize Pattern. Save the answer with -o. Replace VIDEO_ID with the video’s id.

brew install yt-dlp
fabric -y "https://www.youtube.com/watch?v=VIDEO_ID" --pattern summarize -o summary.md

Fabric sends the transcript and the Pattern to your model and writes the reply to summary.md. The YouTube page says --transcript-with-timestamps keeps the time marks. Comments and metadata need a YouTube API key.

For newcomers and experts

If you are new. A Pattern is a prompt saved as a Markdown file, named for the job it does. Fabric adds your text to it and sends both to an AI model. You can try one with nothing installed. The README says you can use any Pattern in any AI application, ChatGPT included. Open the patterns folder on GitHub, copy a Pattern’s text, paste it into your AI app and add your own input. Install the program when you want to run Patterns without the copying.

For experts. Each Pattern is a folder with a system.md file. The README says Patterns use the System section of the prompt almost exclusively, with Markdown structure that spells out what you want and in what order. A strategy adds a prompt change on top. Install them from the fabric -S menu, then pass --strategy with cot, cod, tot, aot, ltm, self-consistent, self-refine, reflexion or standard.

Custom Patterns live in a folder you choose in fabric --setup, under Custom Patterns in the Tools menu. They win over built-in Patterns with the same name, and fabric --updatepatterns never touches them. The README maps a model to a Pattern with an environment variable written as FABRIC_MODEL_PATTERN_NAME=vendor|model, kept in your shell startup files.

The --print-prompt option renders the system and user messages Fabric would compose and stops before any model call, so you can reuse the prompt library with other command-line LLM tools. Some models need raw mode, which merges the two messages. Pass the model with -m to see that structure. Fabric then looks up the vendor, which can send a model-list request.

The server runs only the text and datetime template plugins inside a Pattern. A Pattern that uses a sys, file or fetch plugin, or an extension, gives an error there, while the command line runs all plugins. A loop in your .zshrc or .bashrc turns each Pattern into a command, so summarize replaces fabric --pattern summarize.

Good to know

Fabric sends your input and the Pattern to the provider you chose. These points come from the README and the docs.

• Other connections: the options -u and -q use Jina AI to scrape a page or search a question. The options --serply_search and --firecrawl_search send a search to Serply or Firecrawl. A transcript from -y is fetched with yt-dlp. The option --transcribe-file sends your audio to OpenAI, the only provider it supports
• The server: fabric --serve opens an HTTP server on port 8080. Set --api-key before you bind it to anything but your own computer. The REST API page says the key is an admin key, the server speaks plain HTTP, and GET /config returns the API keys of every configured vendor. It has no rate limiting
• The web app: it has no login. Each client that can open it can use your Fabric server through it, and with --host each computer on the network can
• The installer: the one-line command runs a script from the repository, which downloads the latest release. Read the script first if you want to know what runs
• Docker: the README’s pre-built images come from the kayvan and ksylvan accounts, not the maker’s. Check who builds an image before you run it, or build the maker’s own Dockerfile
• Old videos: the README says many intro videos were recorded when Fabric was Python-based, so follow the install steps here, not the ones in a video
• Updates: for a Go install, the README says to run the same go install command again to get the latest version. It gives no upgrade command for Homebrew, winget or Scoop. Run fabric --updatepatterns for the Patterns. If you run fabric --setup after an upgrade, you may need to enter some settings or keys again. The README says Fabric is evolving rapidly and points to its CHANGELOG
• Removal: the README gives no uninstall steps. Mindket’s notes say what to delete
• License: MIT. The LICENSE file reads “Copyright (c) 2012-2024 Scott Chacon and others”