And Grant gave another good suggestion for starting out bsky.app/profile/gmcd...

5. Ask Claude to use that skill using a "slash command" `/{name-of-skill} perform task x on file @{file-name} at line 123.`

3. Try instructing Claude to create or edit a file to achieve some task.
4. Add a `.claude/skills/{name-of-skill}/SKILL.md` with instructions that you would want yourself or someone else using the code to be aware of when performing a task (potentially the same as above).

So to get back to your question.

1. Start claude from the terminal (or VS Code or Positron or other coding interface) where you already have code written.
2. Run `/init` and claude will create a CLAUDE[dot]md file for you. Read it closely and make sure the file's contents ar accurate.

Then if you have Claude Code open in a terminal, you can write something like `/stata write a do file that creates a summary statistics table of variables x,y,z from auto.dta and export the table to a .tex file` or `use the Stata skill to create a graph of all of the vehicle types in auto.dta`

Then a skill can also include additional instructions
, say in a references/ subfolder of the skill...just more guidance that a human or computer might want to have at their fingertips. I'd say that's the place to start.

In this particular case, ti's instructions that I'd want any RA to be familiar with if I asked them to start work on a particular IPA project. They're guidelines that are based on what we train people to know, and they help give Claude a compass and set the guardrails.

The way I think about Claude Skills is that they are small sets of instructions that you would like a human (or computer) to understand when approaching a task. Have a look a our Stata skill: github.com/PovertyActio....

Yes! Effectively what I did for IPA's Stata Skill, which draws from IPA, World Bank DIME, @seankhiggins.bsky.social, and Gentzkow & Shapiro guides. github.com/PovertyActio...

Has been really beneficial for drafting full Stata end-to-end.

And highly recommend getting comfortable with a little bit of git. It's really helpful to be able to monitor code changes as they accumulate and only committing to a git branch after you've reviewed and tested the code. For a currated set of resources on Git data.poverty-action.org/software/git/

Lasse Brune and I are iterating on the content and structure. Key is getting CLAUDE[dot]md set up for use case and common patterns, use Skills, and provide templates/good examples. We're testing with IPA RAs at the moment and looking for more feedback (general principles would be the same for R.

My team has been using Claude Code extensively with our
python code. We're now rolling out a Stata template that integrates Claude Code for data cleaning and analysis code. github.com/povertyactio...

The sincerest apologies this was stuck in my drafts and I lost my password written on a postit on my desk rejection.

⚾📰 bsky.app/profile/did:...

🔱⚾ bsky.app/profile/did:...

Your park analysis is hands down the best baseball article this year.

Hand made by cross-stitch.