Turn GitHub into local LLMs context

Why GitHub needs preparation before local LLMs

A repository alone is rarely the full project. The AI may also need docs, specs, tickets, or architecture notes that live outside the codebase.

Local models are useful for private workflows, but they usually need tighter bundles because context limits vary by runtime. The result depends heavily on how cleanly your source material is grouped, labeled, and filtered before it reaches the model.

Step-by-step workflow

Add the repository source, include the files relevant to the task, and use filters to remove generated output, dependencies, and test fixtures.

• Add GitHub repositories as a Riflet source.
• Select README files, package manifests, core source files, route handlers, API contracts, schema files, and focused tests.
• Use filters to remove stale exports, generated files, and unrelated background material.
• Export one Markdown or text bundle and use it in local LLMs.

What to include

For this workflow, prioritize README files, package manifests, core source files, route handlers, API contracts, schema files, and focused tests. Add a short instruction file at the top if the AI needs naming conventions, project goals, or constraints.

Do not export the whole repo by default. Include entrypoints, core modules, configuration, docs, and tests that explain the behavior you want the AI to reason about.

Getting the bundle into local LLMs

local LLMs keeps uploaded material in the prompt itself. Paste the bundle into the prompt, or load it through the front-end you run, such as LM Studio, Ollama, or a text-generation web UI. It accepts plain text and Markdown, so a Markdown or plain-text export from Riflet needs no further conversion.

The usable context limit is often far below the number on the model card, because quantisation and available VRAM cap what you can actually load. Check the limit your runtime reports, not the headline figure, before sizing the bundle.

Nothing persists between runs, so the bundle is re-sent every session. That makes bundle size a recurring cost rather than a one-time one, and it is the strongest argument for trimming hard.

Token budget advice

Riflet lets you check the bundle against Local LLM's 32,768-token context window before export. Keep room for your actual question and for the model's answer.

If the bundle is too large, remove files that merely repeat what another file already explains. Context quality usually beats context volume.

Honest limitation

Riflet does not push live updates into local LLMs. It creates the context file. When the source changes, re-export and upload the new bundle yourself.