EasyWrite

Inspiration

We’ve all stared at a blinking cursor, overwhelmed by the daunting task of turning abstract thoughts into a structured document. While LLMs are powerful, they often struggle with long-horizon coherence and maintaining a specific user voice over large documents. We wanted to move beyond simple "autocomplete" and create a system that acts more like a collaborative partner—one that understands context, plans ahead, and helps iterate on ideas rather than just generating text blindly. EasyWrite was born from the desire to bring true agentic capability to the writing process.

What it does

EasyWrite is an intelligent writing agent designed to assist users through the entire lifecycle of content creation. Unlike standard chatbots, it doesn't just output text; it maintains a workspace context. Key features include:

• Recursive Planning: It breaks down complex writing prompts into outlines and sub-tasks before generating content.
• Context Awareness: It remembers previous sections and user constraints to ensure consistency in tone and style.
• Iterative Refinement: Users can ask the agent to rewrite specific sections for clarity or brevity without losing the core message.

How we built it

We built the core of EasyWrite using Python and LangChain.

• The Brain: We utilize Gemini 3 for the reasoning and generation capabilities.
• Memory: To handle long context, we implemente a RAG pipeline to retrieve relevant context dynamically.
• Agentic Workflow: We designed a state machine that allows the agent to switch between "Planning," "Drafting," and "Editing" modes.

Challenges we ran into

The biggest hurdle was context management. Keeping the model focused on the overall narrative arc while working on granular paragraphs was difficult. We also struggled with:

• Latency: Chaining multiple prompts for the "planning" phase initially made the response time too slow.
• Hallucination: Ensuring the agent stuck to the user's provided facts required rigorous prompt engineering and temperature tuning.

Accomplishments that we're proud of

TODO

What we learned

Building EasyWrite taught us that prompt engineering is not enough—system architecture matters most. We learned a great deal about:

• The importance of structured outputs (JSON) for controlling agent behavior.
• Balancing model creativity with strict user constraints.
• How to effectively manage state in a stateless LLM environment.

What's next for EasyWrite

We plan to expand EasyWrite into a full-fledged research assistant. Our roadmap includes:

1. Multi-document ingestion: Allowing the agent to "read" entire folders of background material.
2. ** collaborative mode:** Enabling multiple human users to interact with the same agent instance.
3. Local deployment: Optimizing the agent to run on local hardware for privacy-focused writing.

Built With

• llm
• python