Loose Ends 用本地 AI 從筆記整理待辦清單

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady.

What I Built

Small promises disappear into daily notes: a file to send, a pull request to review, a plan to draft later. Maintaining a second task list means rereading the same notes, which is a low-energy chore I tend to postpone.

Loose Ends scans a folder of Markdown or text notes and produces one reviewable checklist. A local open-weight model recognizes commitments, drops completed items, merges repeated reminders, and proposes a concrete first step. Each item cites the exact note lines. The tool never sends messages or edits the source notes.

The slow-and-steady part is the rerun. When the notes have not changed, a SHA-256 cache returns the last report with zero model calls. I can run it each morning without spending inference on the same text.

Demo

Watch the 16-second deployed demo. It replays a verified local run using two synthetic notes. Monday says I promised Mina an API schema and should review a pull request. Tuesday repeats the promise and adds a later demo outline. The model produced three checklist items, merged the duplicate, and ignored a completed hosting invoice.

First run: Wrote 3 review items to demo-report.md
Second run: Unchanged notes; reused demo-report.md (no model call)

Sample input · Actual output

Code

Source, setup instructions, and MIT license. The Python CLI uses only the standard library and a local Ollama endpoint. No cloud key is needed.

How I Built It

The reader walks .md and .txt files, labels each line filename:line, and limits file counts and sizes. The prompt asks a local qwen2.5-coder:7b-instruct-q4_K_M model for structured JSON: task, urgency, next step, reason, and source lines. Validation discards any suggestion that lacks a real source token. The renderer groups items by today, this_week, later, or unknown.

I tested the actual Ollama run on an RTX 4060 laptop. The sample run returned three tasks; a second run reused the report. The first call took about twelve seconds in this run. Performance will vary by model and hardware.

This is intentionally review-first. Automatic mail sending or file moves would turn a model error into a real-world mistake. The tool gives me a compact checklist and evidence so I can decide what to do.

Why Does Open Innovation Matter?

Notes can contain private work and names. Running an open-weight model through local Ollama keeps the note text on the machine rather than sending it to a hosted model API. I can swap models, inspect the prompt and validation, and rerun at no per-call charge. That makes a small recurring chore worth automating.

The Lazy Payoff

The sample run turned two notes into three review items and merged one repeated promise. The next run needed no inference. I still check each cited line before acting: a local model can be wrong, and the evidence is part of the product.