Show HN: I built an AI that turns GitHub codebases into easy tutorials

Original Article ↗ Hacker News Discussion ↗

Project concept & overall reception

  • Tool turns GitHub repos into multi-chapter tutorials with diagrams using LLMs (primarily Gemini 2.5 Pro).
  • Many commenters are impressed, calling it one of the more practical and compelling AI applications they’ve seen, especially for onboarding and understanding unfamiliar libraries.
  • Some tried it on their own or employer codebases and reported surprisingly accurate, useful overviews with minimal manual edits.

Capabilities, models, and architecture

  • Uses Gemini 2.5 Pro’s 1M-token context and strong code reasoning; designed explicitly around these new “reasoning” models.
  • No classic RAG pipeline; instead feeds large swaths of code directly and orchestrates multi-step prompting, documented in a design doc.
  • Supports swapping to other models (OpenAI, local via Ollama), though quality is reported as lower with smaller/local models.
  • Repo/file selection is regex‑based and currently excludes tests/docs by default; several people question that design choice.

Large and complex codebases

  • Linux-size repositories exceed context limits; suggested approaches:
    • Decompose into modules (kernel vs drivers, per-architecture, AST-based partitions).
    • Wait for even larger context models.
  • Contributors to major projects (e.g., OpenZFS, LLVM) outline desired outputs: subcomponent overviews, disk formats/specs, advanced feature internals, plugin architectures, optimization pass guides.

Tone, style, and tutorial quality

  • Major recurring criticism: writing is over-cheerful, full of exclamation marks and “cute” analogies that feel patronizing or vacuous to engineers.
  • Others argue beginner-friendly, analogy-heavy tone has value for non-experts or PMs.
  • The style is prompt-driven and can be edited in the code; multiple prompt suggestions are shared to make text more rigorous and less ELI5.
  • Some say content can drift into generic theory (e.g., long explanations of “what an API is”) rather than action-oriented tutorials.

Use cases: onboarding & documentation maintenance

  • Strong interest in:
    • Onboarding to large existing systems (databases, OS kernels, enterprise frameworks).
    • Continuous documentation maintenance: using diffs/commits to update docs, or having the tool flag mismatches between code and docs.
    • Generating missing high-level architecture docs and “how to use” guides based on tests and usage examples.
  • A technical writer notes this could expand, not replace, demand for human docs work by making high-quality docs more feasible, shifting humans into orchestration and review.

AI usefulness, limits, and hype debate

  • Many see this as a concrete rebuttal to “AI is pure hype,” especially for code comprehension and summarization.
  • Others caution that:
    • LLMs still hallucinate, especially on mature, messy, business-logic-heavy codebases.
    • Tools can mislead if their confident summaries are taken as ground truth.
    • True “why” documentation still requires human intent and context.
  • Debate over claims like “built an AI”: some view this as overstated marketing for what is essentially a sophisticated LLM-powered app.

Practicalities: cost, setup, and reliability

  • Reported cost example: ~4 tutorials for about $5 on Gemini API.
  • Free tiers (e.g., Gemini’s daily request limits) let users experiment on a few repos.
  • Some note Gemini 2.5 Pro is still “preview” and can be flaky; others prefer alternative models.
  • Several users discuss adding CI/CD or GitHub Actions integration, private repo access via tokens or local directories, and potential extension into interactive or guided usage tutorials.