SlothDesk：別再翻閲客戶對話，改為制定清晰工作計劃

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

A client asks you to change a heading. Ten messages later, they cancel that request and ask you to fix the contact form instead. The next morning, you read the entire thread again to figure out what actually needs doing.

I built SlothDesk to prepare that work plan for review.

What I Built

SlothDesk turns client conversations into proposed tasks, decisions, clarification questions and a reply draft. Paste a conversation or upload a text export, inspect the supporting messages, and accept the changes you want on your project board.

The most useful part is handling changed instructions: the final work plan should reflect what the client currently wants. Cancelled requests, revised deadlines and follow-up messages deserve more than another summary.

Demo

Try SlothDesk

Create a project, load a bundled English or Hinglish conversation, and analyse it. Inspect the proposed changes and source messages before accepting selected items. Then import a follow-up to review how the plan changes.

Hosted demo limitation: the deployed app currently reports that Ollama is unavailable and uses its rule-based fallback. This demonstrates the workflow, not an actual open-weight model run. The repository provides the local Ollama integration.

Code

🦥 SlothDesk — Skip the chat scroll. Get the work.

SlothDesk turns messy client conversations into an actionable work plan using AI running locally on your computer. Paste a chat (or upload a .txt export) let it analyse, review the proposed changes, and accept the ones you want.

Built for freelancers, agency owners, and project managers who otherwise re-read long client threads to reconstruct tasks, deadlines, and changed instructions by hand.

Works in mixed Hindi/English (Hinglish) and English.

The one workflow it does well

Import chat  →  Review proposed changes  →  Accept  →  Project board
   (paste/upload)   (New/Updated/Cancelled/      (commit     (edit, status,
                     Needs-review + evidence)     selected)    follow-up, export)

The model’s output is never written straight to your board. Every analysis produces a proposal you review and accept — so you stay in control and nothing is invented into your task list.

Its signature trick is reconciliation: when a client says…

How I Built It

The project uses Next.js, TypeScript and Tailwind. The local inference path calls Ollama with a configurable model, defaulting to Gemma 3 4B. JSON Schema and Zod validate the extraction structure; application logic handles message IDs, date resolution, reconciliation and persistence.

Model output becomes a proposal. It does not automatically become an accepted task. That review step is central to the product.

The local configuration uses SQLite. According to the README, the serverless deployment falls back to ephemeral in-memory storage, so the hosted demo should not be treated as durable project storage.

Why Open Innovation Matters

With the local setup, users can run an open-weight model on hardware they control, inspect the extraction prompt, and replace the model. They do not need to build this workflow around a single proprietary inference provider.

Gemma weights have their own terms; the application code is MIT licensed. Local inference avoids per-request hosted inference fees, while still requiring hardware and electricity.

Privacy claims apply to the local configuration. The hosted website is a server-backed demo: do not upload confidential conversations on the assumption that processing stays in your browser.

Evaluation and Limitations

The README reports 14 passing cases for the deterministic fallback on its small bundled dataset. These are author-reported fallback results, not a model-quality benchmark or evidence of performance on arbitrary client conversations.

Actual local-model performance and time savings still need measured validation. I am not claiming a percentage of time saved.

The current scope is text input, proposed work plans, source evidence and human review. Live WhatsApp/Slack integrations, automatic sending and voice transcription are outside this version.

The Lazy Payoff

The intended payoff is straightforward: stop reconstructing a task list from the same conversation repeatedly. Let the software prepare a reviewable plan, then use your attention on the decisions that still need a person.

AI disclosure: This article and cover were generated with AI tools from my project materials and instructions. Model benchmarks and time savings have not been independently verified.