Helping yourself

2010s

Let me GitHub search that for you

2020s

https://github.com/search?type=code

Let me ask an LLM for you

2026

https://claude.ai · Posit Assistant

Effective self-help tips

  • Being 20% more specific improves your results by 80%
  • Ways to introduce specificity
    • Your own language
    • Narrow the scope, e.g., to a specific codebase
    • Use a tool with unique access to, e.g., a specific codebase
    • Commision a script to do something exhaustively
  • Don’t be shy to ask for receipts! E.g. show me where this appears in the codebase or the docs

Talk like an expert

Terence Tao discussing a surprising math result with ChatGPT

It is remarkable that the jacobian is constant, that is an exceptional amount of cancellation. Does this polynomial map have any symmetry or other structure that makes this cancelation less miraculous?

What stands out is the shape of the interaction. Tao starts his inquiry by noting a single, precise detail that intrigues him.

Code is the byproduct — yagmin.com

Terse and precise

Note what Tao is not doing.

He is not building a rich contextual background for the LLM to unpack. He does not list his credentials or say, “you are a mathematical genius, make no mistakes.”

Instead, he is terse and precise. He asks narrow questions and gets narrow responses.

Keep your requests narrow and specific. Let the output compound through repeated inquiries. Engage with the goal of improving your personal understanding.

Code is the byproduct — yagmin.com

Chat: where can an R package write user files?

GitHub search: Show usage of tools::R_user_dir()
in CRAN packages

Run this search

DeepWiki: What does pkgdown use tools::R_user_dir() for?

https://deepwiki.com/r-lib/pkgdown

DeepWiki: What does pkgdown use tools::R_user_dir() for?

GitHub Chat with Copilot: When did pkgdown introduce tools::R_user_dir()?

Visit https://github.com/r-lib/pkgdown and click the Chat with Copilot button

GitHub Chat with Copilot: When did pkgdown introduce tools::R_user_dir()?

Visit https://github.com/r-lib/pkgdown and click the Chat with Copilot button

Agent: Create a Quarto document surveying tools::R_user_dir() usage

OK I’m not actually doing this one, but you could imagine directing an agent to:

  • Enumerate all uses of tools::R_user_dir() in CRAN packages
  • Create a raw report of GitHub permalinks to these call sites
  • Analyze each call site re: how it’s being used and when it was introduced
  • Summarize the findings

It’s a great time to answer your own questions!

  • Chat: be as precise as your current understanding allows
  • GitHub search: qualifiers are powerful, e.g.
    • user:cran searches across source of every CRAN package
    • user:tidyverse searches within the tidyverse organization
    • path:*.yaml search within YAML files
    • repo:wch/r-source searches the GitHub mirror of R’s source
    • language:R searches R code
  • GitHub Chat with Copilot: knows about commits, issues, pull requests, in addition to code
  • DeepWiki: chat on top of an LLM-generated wiki for a codebase

Code is the byproduct of understanding

How do you construct shared understanding? By narrowing the LLM’s focus through domain-specific wording and keeping requests specific and cumulative.

Tao pins ChatGPT to a “deep mathematics” headspace through terminology and the specificity of his question.

If you don’t understand something, set the stage for the LLM and it will capably fill in the missing 20%.

Code is the byproduct — yagmin.com