Retype agent-written diffs to internalize new features
After an agent writes a feature, a skill shows the important diffs against a copy of the base branch and the author retypes them to understand the code. ·
Done with a custom-built or unnamed agent
- What they did
- After an agent writes a new feature, fcoury runs a skill that makes a copy of the base branch and presents each diff that matters, skipping imports and other boilerplate. He then types each change into an editor himself. He has also tried asking questions about specific aspects and diffs with gaps to fill in.
- How it went
- Retyping works well for critical work, but the time it takes means he uses it sparingly. The question and fill-in-the-gap variants help somewhat, and he is still looking for better methods and research.
- Worth knowing
- It is slow, so reserve it for critical features rather than everyday changes.
Try it yourself with Claude Code
I just had you build [feature name] on the branch [branch name]. Compare it against a copy of [base branch] and show me only the most important diffs, one chunk at a time, with a one-sentence explanation of why each matters. Don't modify any files; I'll retype each chunk myself, and when I say 'next' show me the following chunk until we've covered everything essential.
Want help getting this working? Join the Vibe Coding Academy ↗ Run by Alex Finn, who also runs this site.
Source: HN · Nov 8, 2026
Related use cases
Build a C compiler with a team of parallel agents
Anthropic's engineering write-up on tasking Opus 4.6 agent teams with building a C compiler.
Build a full web app with database, storage and landing page
The author used Replit Agent to build a Postgres-backed web app with image storage, analytics and a landing page; an S3 switch failed and was reverted.
Build custom real estate tools for offers and deal tracking
A rental owner used Codex to build tools for formulating offer prices, building property analysis reports with field photos, and tracking transactions under contract.
Control a browser to fill forms and debug sites
A tutorial describes Scout opening sites, filling forms, clicking, uploading files, extracting page info, and reading console logs and network requests.
Scan every git commit for secrets and injection flaws
A free tool runs 11 Claude Code CLI agents on each commit to catch secrets, injection flaws, and bad migrations before code lands.
Build a formally verified Lean web app runtime with an agent
An agentic engineer had Claude build Lasm, which runs Lean web apps in browser, Node, serverless and edge environments, under his high-level direction.
Get 5 like this every morning
The best things people got an AI agent to do, each with the prompt to try it.