IdeaWell

Project ideas from Hacker News discussions.

Redesigning source control at Anthropic

Original Article

Hacker News Discussion

📝 Discussion Summary (Click to expand)

Four Prevalent Themes in the Hacker News Discussion on Anthropic's Scion

  1. Prompts vs. Specs & Source Control Paradigm Shift: Debate centers on whether AI prompts can function as reliable "source code" or specifications, challenging traditional version control assumptions.
    > amelius: It would be nice if you could just commit the prompts as your "source code", and let the AI-generated code sit in the repo as a kind of expensive-to-generate build output.
    > 6thbit: Prompts are not specs.
    > spuz: You can do that - it's called Spec Driven Development.

  2. Skepticism About Scion's Utility & Motives: Widespread doubt regarding Scion's broader applicability, reproducibility, and whether it solves real problems or serves Anthropic's internal hype/IPO goals.
    > amluto: In Scion, every commit, rebase, or merge happens on the server... I think this model sucks even in a pervasive AI world... But don’t lose the actual benefits of git!
    > simianwords: Anthorpic invanted new version control to do hype marketing and cash in the IPO before the bubble busrts (true story).
    > GuB-42: Do we have reproducible LLMs? Possible in theory, with a seeded PRNG, but in practice?

  3. Scale of AI-Driven Development Velocity: Astonishment and scrutiny over the unprecedented commit rates (30k+/day), questioning what justifies such volume and its implications for engineering value.
    > smy20011: as of September, we merged over 30,000 per day.
    > temac: 30000 a day seems already beyond absurd, it's one merge every 2.9s... winfredJa: im still not sure what they built software wise that warranted these many commits.
    > piker: Honest question: at this point, what do people do at Anthropic? It sounds like Claude loops on itself and must be pushing millions of lines of code changes per day.

  4. Evolving Role of Human Engineers: Shift toward humans providing high-level intent, oversight ("saying no"), and architectural guidance, while AI handles low-level code generation and execution.
    > tetraodonpuffer: Our engineers still supplied the invaluable understanding of the high-level inter-system components... Very often, we found that the role of the engineers was to tell Claude “no.”
    > jareds: Is this really something to be proud of? [...] Is the future of our industry going to be oneshotting prompts with out manual code review and then having AI fix the bugs found in production buy customers?
    > CaptWorld: They review based on the intent of the change not jz blindly.. the job of a developer is to provide high level design decisions and saying no to claude so that it doesn't over engineer solutions.

🚀 Project Ideas

Generating project ideas…

PromptSpec

Summary

  • Store LLM prompts as version-controlled source files and treat AI-generated code as reproducible build outputs.
  • Enables deterministic builds via seeded RNG and prompt hashing, solving the "prompts are not specs" frustration.

Details

Key Value
Target Audience AI-assisted developers, teams using LLM coding agents
Core Feature CLI that hashes prompts, runs LLM with fixed seed, caches output, and records prompts as source
Tech Stack Rust (for speed), Python bindings, SQLite cache, optional WASM
Difficulty Medium
Monetization Hobby

Notes

  • HN commenters lamented: "Prompts are not specs." (6thbit) and desire to commit prompts as source code (amelius).
  • Provides a concrete workflow for spec-driven development with reproducible AI output, useful for discussion on versioning AI-generated code.

ScionGit Bridge

Summary

  • Offline-first compatibility layer that lets developers work with a Scion-like VCS locally using git, then sync to a server-side Scion backend.
  • Addresses the pain of online-only VCS and the need to migrate existing git workflows.

Details

Key Value
Target Audience Enterprises evaluating AI-native VCS, teams wanting to try Scion without abandoning git
Core Feature Bidirectional sync: local git repo ↔ remote Scion server, with conflict resolution
- Tech Stack Go (for server), libgit2 bindings, gRPC, Protobuf
Difficulty High
Monetization Revenue-ready: SaaS tiered pricing (per active user)

Notes

  • Users complained about Scion requiring server for every commit: "In Scion, every commit, rebase, or merge happens on the server" (amluto) and desire for offline work.
  • Bridge eases adoption, inviting discussion on hybrid VCS models and reducing lock‑in concerns.

DeterministicLLM

Summary

  • Wrapper service that ensures LLM calls are reproducible by fixing seeds, controlling temperature, and caching results by prompt+seed.
  • Solves the reproducibility challenge highlighted by GuB-42 and batshit_beaver.

Details

Key Value
Target Audience Teams using LLMs for code generation, build systems, CI pipelines
Core Feature API proxy that returns deterministic outputs, logs seed, and provides cache invalidation
Tech Stack Node.js/TS, Redis cache, OpenAPI spec, Docker
Difficulty Low
Monetization Hobby

Notes

  • Commenters asked: "Do we have reproducible LLMs? Possible in theory, with a seeded PRNG, but in practice?" (GuB-42).
  • Provides a practical solution that lets teams treat LLM output like compiled artifacts, enabling reliable version control and diffing.

IntentReviewer

Summary

  • AI-powered code review bot that checks whether a PR's implementation matches its description/intent, using LLMs to summarize code and compare against PR text.
  • Addresses the bottleneck of reviewing thousands of agent‑generated PRs per day (tetraodonpuffer).

Details

Key Value
Target Audience Organizations with high‑volume PR streams, especially those using AI coding agents
Core Feature Automated intent‑matching report: highlights mismatches, suggests edits, and scores confidence
Tech Stack Python, FastAPI, LLM API (Anthropic/OpenAI), vector store for code embeddings
Difficulty Medium
Monetization Revenue-ready: per‑repo monthly fee

Notes

  • tetraodonpuffer noted: "I am not sure however how easy it is to review 30k PRs/day even if only for intent..." indicating need for automated intent checks.
  • Would spark discussion on balancing AI review with human oversight and improve trust in agent‑generated code.

Read Later