# Sloth Brief — let a local model read the notes you keep postponing

> Source: <https://dev.to/joly_bigj/sloth-brief-let-a-local-model-read-the-notes-you-keep-postponing-2jo4>
> Published: 2026-10-23 10:00:42+00:00

Sloth Brief turns a handful of accumulated text notes into a short, source-linked action brief. It is a small contribution to the “Slow and Steady” theme: spend less effort rereading notes when restarting a task.

The verified runtime is local Hugging Face Transformers with Qwen2.5-0.5B-Instruct. Ollama is an optional adapter, not the runtime used for the demonstrated result. There is no paid API, account requirement for inference, or cloud upload of notes. The first run downloads model weights; subsequent inference uses the local cache.

[Source code and installation instructions](https://github.com/1adwad-BigJ/sloth-brief)

Run the reproducible demo:

```
pip install -r requirements.txt
python demo.py
python -m unittest -v
```

The demo contains two synthetic notes. One says the documentation draft needs its installation section reviewed before publication. The other describes a CSV importer failure on empty files and asks for a regression test.

Actual CPU output on October 23, 2026:

Review the installation section before publishing. [sources: 1]

Reproduce CSV importer failure with empty file fixture; add regression test; wait for fix. [sources: 2]

[Generated Markdown](https://github.com/1adwad-BigJ/sloth-brief/blob/main/demo-verified.md) · [Model responses](https://github.com/1adwad-BigJ/sloth-brief/blob/main/demo-model-response.json) · [Reproducible demo script](https://github.com/1adwad-BigJ/sloth-brief/blob/main/demo.py)

Eight automated tests pass. These test input limits, malformed results, source references, source hashes and read-only file handling; they do not establish model accuracy on arbitrary notes.

[Qwen2.5-0.5B-Instruct](https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct) provides openly available Apache-2.0 weights. Transformers and PyTorch run inference on the user's machine. Users can inspect the application, replace the model and process notes without paying an inference provider. Sloth Brief's code is MIT licensed.

This is deliberately a small CPU-capable build. It summarizes at most five notes per Transformers run, separately, rather than trying to operate as an unrestricted agent.

The initial version asked the small model for a JSON action list. A real run produced duplicate keys and invalid source references. Instead of claiming that valid-looking JSON solved grounding, the final version asks for one concise summary per note. Python assigns the source number from the note actually supplied to the model.

That makes source attribution deterministic, but it does not guarantee that a summary is faithful. The reader should still check the original note. The model can misunderstand instructions or omit details; the demo is a verified example, not a broad benchmark.

Input files are never edited. The output must be a new file. Size limits bound the workload, and source hashes help identify the note version used. The model has no tools or execution privileges. A malicious note can influence its summary, but cannot cause it to run shell commands through this application.

The payoff is a short restart checklist instead of another pass through every note. There is no measured productivity claim: this prototype demonstrates the workflow on two synthetic examples.

Overall prize. No partner category is claimed.

Built with Codex assistance during the challenge window. Code and writing were prepared with AI assistance. The model demo ran locally on real open weights; it is not a mocked response. No client data was used.