The four parts of a useful context bundle
After working with hundreds of context bundles across different domains, a clear pattern emerges. The most effective bundles contain four types of material, each serving a distinct purpose.
Project brief
A concise overview that tells the AI what the project is, what stage it is in, and what the key constraints are. This is the single most important element. Without it, the AI is guessing about the basics. A good project brief is one to two paragraphs long and answers: what is this? what are we trying to achieve? what decisions have already been made?
Reference documents
The actual source material the AI needs to work with: research notes, specifications, past work, meeting summaries, data, or design documents. These give the AI factual grounding so it draws from your real information instead of inventing plausible-sounding content.
Style and tone guidance
How should the AI write? What voice, vocabulary, sentence structure, or formatting conventions should it follow? This is especially important for writing projects, brand content, and code. Include a style guide, example paragraphs, or a list of tone adjectives so the AI matches your standards.
AI instructions
Explicit instructions for the AI: what kind of help you want, what to avoid, what format to use for responses, and any guardrails. For example: "Always cite the source document when making claims" or "Do not introduce new characters that are not in the character sheet."
How large should a context bundle be
The ideal size depends on your model. ChatGPT (GPT-4o) has a 128,000-token context window, which fits about 96,000 words. Claude offers 200,000 tokens, roughly 150,000 words. In practice, you want to leave at least 20 to 30 percent of the window free for the conversation itself.
A good target is to keep your bundle under 50,000 words. This fits comfortably in both ChatGPT and Claude, and it is small enough that the AI does not lose focus across the content. Quality over quantity: 10 well-chosen documents outperform 50 loosely related ones.
Rule of thumb: If a human colleague would need a piece of information to do the job well, include it in the bundle. If they would not need it, leave it out. Treat the bundle like a briefing document, not an archive.
How often to update your bundle
Update your context bundle at the start of a new phase or when core facts change, not after every conversation. A good bundle represents the stable knowledge about a project. Most users have one bundle per active project and update it every week or two.
The goal is that the AI immediately understands your project the moment you paste the bundle, without you needing to add any explanation. If you find yourself typing clarifications at the start of each session, your bundle needs updating.
Best format for context bundles
Markdown is the most effective format. AI models like Claude and ChatGPT handle Markdown well: they understand headers, bullet points, and code blocks, which helps them navigate the context quickly. Use clear section headers to separate each source file or category. Include filenames as headers so the AI can reference specific sources.
Riflet exports in Markdown by default, with each source file clearly labeled and separated. The output is immediately ready to paste into any AI chat.
Build your first context bundle
Riflet helps you select, organize, and export your project files into one AI-ready bundle.
Download Riflet free