Show HN: Runprompt – run .prompt files from the command line

Posted by chr15m 12 days ago

I built a single-file Python script that lets you run LLM prompts from the command line with templating, structured outputs, and the ability to chain prompts together.

When I discovered Google's Dotprompt format (frontmatter + Handlebars templates), I realized it was perfect for something I'd been wanting: treating prompts as first-class programs you can pipe together Unix-style. Google uses Dotprompt in Firebase Genkit and I wanted something simpler - just run a .prompt file directly on the command line.

Here's what it looks like:

--- model: anthropic/claude-sonnet-4-20250514 output: format: json schema: sentiment: string, positive/negative/neutral confidence: number, 0-1 score --- Analyze the sentiment of: {{STDIN}}

Running it:

cat reviews.txt | ./runprompt sentiment.prompt | jq '.sentiment'

The things I think are interesting:

* Structured output schemas: Define JSON schemas in the frontmatter using a simple `field: type, description` syntax. The LLM reliably returns valid JSON you can pipe to other tools.

* Prompt chaining: Pipe JSON output from one prompt as template variables into the next. This makes it easy to build multi-step agentic workflows as simple shell pipelines.

* Zero dependencies: It's a single Python file that uses only stdlib. Just curl it down and run it.

* Provider agnostic: Works with Anthropic, OpenAI, Google AI, and OpenRouter (which gives you access to dozens of models through one API key).

You can use it to automate things like extracting structured data from unstructured text, generating reports from logs, and building small agentic workflows without spinning up a whole framework.

Would love your feedback, and PRs are most welcome!

Comments

Comment by Barathkanna 12 days ago

This is really clever. Dotprompt as a thin, pipe-friendly layer around LLMs feels way more ergonomic than spinning up a whole agent stack. The single-file + stdlib approach is a nice touch too. How robust is the JSON schema enforcement when chaining multiple steps?

Comment by chr15m 12 days ago

If the LLM returns invalid schema the next link in the chain will send that through since it defaults to string input if the JSON doesn't parse. Maybe I should make it error out in that situation.

Comment by anonym29 12 days ago

Is including a json schema validator and running the output through the validator against the input prompt, such that you can detect when the output doesn't match the schema, and optionally retry until it does match (possibly with a max number of attempts before it throws an error) too complex of an idea for the target implementation concept you were envisioning?

It certainly doesn't intuitively sound like it matches the "Do one thing" part of the Unix philosophy, but it does seem to match the "and do it well" part.

That said, I can totally understand a counterargument which proposes that schema validation and processing logic should be something else that someone desiring reliability pipes the output into.

Comment by chr15m 11 days ago

I'm not sure. I think I need to use it more to see what the LLMs do with bad data. The design you're suggesting might be the answer though.

Comment by threecheese 11 days ago

Looks like Google has packaged dotprompt into a Python library, might allow you to make the codebase leaner: https://github.com/google/dotprompt/tree/main/python/dotprom...

I think you mentioned elsewhere that you dont want to have a lot of dependencies, but as the format evolves using the reference impl will allow you to work on real features.

Comment by chr15m 11 days ago

Will have a look at this. That could be the way to go. Thanks.

Comment by cootsnuck 12 days ago

This is pretty cool. I like using snippets to run little scripts I have in the terminal (I use Alfred a lot on macOS). And right now I just manually do LLM requests in the scripts if needed, but I'd actually rather have a small library of prompts and then be able to pipe inputs and outputs between different scripts. This seems pretty perfect for that.

I wasn't aware of the whole ".prompt" format, but it makes a lot of sense.

Very neat. These are the kinds of tools I love to see. Functional and useful, not trying to be "the next big thing".

Comment by PythonicNinja 12 days ago

added some examples using runprompt in blog post:

"Chain Prompts Like Unix Tools with Dotprompt"

https://pythonic.ninja/blog/2025-11-27-dotprompt-unix-pipes/

Comment by chr15m 12 days ago

Great article, thanks.

"One-liner code review from staged changes" - love this example.

Comment by dymk 12 days ago

Can the base URL be overridden so I can point it at eg Ollama or any other OpenAI compatible endpoint? I’d love to use this with local LLMs, for the speed and privacy boost.

Comment by jedbrooke 12 days ago

https://github.com/chr15m/runprompt/blob/main/runprompt#L9

seems like it would be, just swap the openai url here or add a new one

Comment by chr15m 11 days ago

I've implemented this now. You can set it with BASE_URL or OPENAI_BASE_URL which seems to vaguely be the standard. I also plan to use this with local LLMs. Thanks for the suggestion!

Comment by chr15m 12 days ago

Good idea. Will figure out a way to do this.

Comment by benatkin 12 days ago

Perhaps instead of writing an llm abstraction layer, you could use a lightweight one, such as @simonw's llm.

Comment by chr15m 12 days ago

I don't want to introduce a dependency. Simon's tool is great but I don't like the way it stores template state. I want my state in a file in my working folder.

Comment by threecheese 11 days ago

Can you explain this decision a bit more? I’m using ‘llm’ and I find your project interesting.

Comment by chr15m 11 days ago

llm stores data (prompts, responses, chats, fragments, aliases, attachment metadta) in a central sqlite database outside the working directory, and you have to use the tool to view and manipulate that data. I prefer a tool like this to default to storing things in a file or files in the project directory I'm working in, in a way that is legible e.g. plain text files. Contrast with e.g. git where everything goes into .git.

Functions require you to specify them on the command line every time they're invoked. I would prefer a tool like this to default to reading the functions from a hierarchy where it reads e.g. .llm-functions in the current folder, then ~/.config/llm-functions or something like that.

In general I found myself baffled when trying to figure out where and how to configure things. That's probably me being impatient but I have found other tools to have more straightforward setup and less indirection.

Basically I like things to be less centralized, magic, and less controlled by the tool.

Another thing, which is not the fault of llm at all, is I find Python based tools annoying to install. I have to remember the env where I set them up. Contrast with a golang application which is generally a single file I can put in ~/bin. That's the reason I don't want to introduce a dep to runprompt if I can avoid it.

The final thing that I found frustrating was the name 'llm' which makes it difficult to conduct searches as it is the generic name for what the thing is.

It is an amazing piece of engineering and I am a huge fan of simonw's work, but I don't use llm much for these reasons.

Comment by khimaros 12 days ago

simple solution: honor OPENAI_API_BASE env var

Comment by tomComb 12 days ago

Everything seems to be about agents. Glad to see a post about enabling simple workflows!

Comment by oddrationale 12 days ago

Interesting! Seems there is a very similar format by Microsoft called `.prompty`. Maybe I'll work on a PR to support either `.prompt` or `.prompty` files.

https://microsoft.github.io/promptflow/how-to-guides/develop...

Comment by chr15m 12 days ago

Oh interesting. Will investigate, thanks!

Comment by gessha 12 days ago

Just like Linus being content with other people working on solutions to common problems, I’m so happy that you made this! I’ve had this idea for a long time but haven’t had the time to work on it.

Comment by __MatrixMan__ 12 days ago

It would be cool if there were some cache (invalidated by hand, potentially distributed across many users) so we could get consistent results while iterating on the later stages of the pipeline.

Comment by stephenlf 12 days ago

That’s a great idea. Store inputs/outputs in XDG_CACHE_DIR/runprompt.sqlite

Comment by chr15m 12 days ago

Do you mean you want responses cached to e.g. a file based on the inputs?

Comment by __MatrixMan__ 12 days ago

Yeah, if it's a novel prompt, by all means send it to the model, but if its the same prompt as 30s ago, just immediately give me the same response I got 30s ago.

That's typically how we expect bash pipelines to work, right?

Comment by chr15m 11 days ago

Bash pipelines don't do any caching and will execute fresh each time, but I understand your idea and why a cache is useful when iterating on the command line. I'll implement it. Thanks!

Comment by chr15m 10 days ago

I've now added caching to runprompt with the --cache flag and RUNPROMPT_CACHE env var. Thanks for the suggestion!

Comment by dymk 12 days ago

“tee” where you want to intercept and cat that file into later stages?

Comment by __MatrixMan__ 12 days ago

Yeah sure but it breaks the flow that makes bash pipelines so fun:

- arrow up

- append a stage to the pipeline

- repeat until output is as desired

If you're gonna write to some named location and later read from it you're drifting towards a different mode of usage where you might as well write a python script.

Comment by 12 days ago

Comment by meander_water 12 days ago

This is really cool and interesting timing, as I created something similar recently - https://github.com/julio-mcdulio/pmp

I've been using mlflow to store my prompts, but wanted something lightweight on the cli to version and manage prompts. I setup pmp so you can have different storage backends (file, sqlite, mlflow etc.).

I wasn't aware of dotprompt, I might build that in too.

Comment by cedws 12 days ago

Can it be made to be directly executable with a shebang line?

Comment by leobuskin 12 days ago

/usr/local/bin/promptrun

  #!/bin/bash
  file="$1"
  model=$(sed -n '2p' "$file" | sed 's/^# \*//')
  prompt=$(tail -n +3 "$file")
  curl -s https://api.anthropic.com/v1/messages \
    -H "x-api-key: $ANTHROPIC_API_KEY" \
    -H "content-type: application/json" \
    -H "anthropic-version: 2023-06-01" \
    -d "{
      \"model\": \"$model\",
      \"max_tokens\": 1024,
      \"messages\": [{\"role\": \"user\", \"content\": $(echo "$prompt" | jq -Rs .)}]
    }" | jq -r '.content[0].text'

hello.prompt

  #!/usr/local/bin/promptrun
  # claude-sonnet-4-20250514

  Write a haiku about terminal commands.

Comment by chr15m 10 days ago

This is cool. An even more minimal bash version. Love it!

Comment by 12 days ago

Comment by _joel 12 days ago

it already has one - https://github.com/chr15m/runprompt/blob/main/runprompt#L1

If you curl/wget a script, you still need to chmod +x it. Git doesn't have this issue as it retains the file metadata.

Comment by vidarh 12 days ago

I'm assuming the intent was to as if the *.prompt files could have a shebang line.

   #!/bin/env runprompt
   ---
   .frontmatter...
   ---
   
   The prompt.

Would be a lot nicer, as then you can just +x the prompt file itself.

Comment by chr15m 11 days ago

I added this and you can now make .prompt files with a runprompt shebang.

#/usr/bin/env runprompt

Comment by chr15m 12 days ago

That's on my TODO list for tomorrow, thanks!

Comment by ltbarcly3 12 days ago

Ooof, I guess vibecoding is only as good as the vibecoder.

Comment by stephenlf 12 days ago

Seeing lots of good ideas in this thread. I am taking the liberty of adding them as GH issues

Comment by stephenlf 12 days ago

Fun! I love the idea of throwing LLM calls in a bash pipe

Comment by journal 12 days ago

i literally vibe coded a tool like this. it supports image in, audio out, and archiving.

Comment by chr15m 12 days ago

Cool, I'm going to add file modalities too. Thanks for the validation!

Comment by orliesaurus 12 days ago

Why this over md files I already make and can be read by any agent CLI ( Claude, Gemini, codex, etc)?

Comment by jsdwarf 12 days ago

Claude.md is an input to claude code which requires a monthly plan subscription north of 15€ / month. Same applies to Gemini.md, unless you are ok that they use your prompts for training Gemini. The python script works with a pay per use api key.

Comment by chr15m 12 days ago

Less typing. More control over chaining prompts together. Reproducibility. Running different prompts on different providers and models. Easy to install and runs everywhere. Inserts into scripting workflows simply. 12 factor env config.