Initial Pass at documents necessary for AI assisted programming

[Zelos Zhu (zdz2101)]

Overview

Working in the gsm ecosystem of packages, it can prove difficult attaching the appropriate and full extent of context when using coding assistants, and even if you make some .md's , it can easily "drift" from the purpose/north star of our gsm ecosystem of packages.

With the powers of GPT-5.3-Codex, I've prototyped a slew of appropriate .mds to "synchronize" a context window across our ecosystem and built 2 functions that can rapidly improve/speed up initial prompting to copilot (the tool Gilead is already providing for us) of any gsm ecosystem issue

• Under inst/ai_templates, I'll just call them a family of docs that pretty much describe basic development practices, etc. And maybe they're not all entirely necessary but they were needed/artifacts of spawning this idea/framework. Specifically I want to highlight ECOSYSTEM.md where we can enrich with examples or thorough descriptions of each package across all repos and a templated ARCHITECTURE.md which will be repo specific. I also added a new issue template denoted a 6-CONTEXT_PACK.md that feels like the necessary context alongside the ai docs to one-shot prompt copilot, for a pretty good first pass at code generation. I think if we iterate on the ai docs and then this refined context pack issue template it'll get better:

• 2 functions were built:

1. sync_gsm_standards() which ensures our issue templates/cicd/ai docs stay synced across repos, and an associated cicd check for this has been created as well via ai-template-drift-check.yaml to especially ensure agents/coding assistants don't go too haywire
2. build_agent_prompt() will create the appropriate templated prompt to grab & read these ai docs and begin a first pass at issues, it is as simple as build_agent_prompt("gsm.qtl#78") , copy/paste into the copilot chat on VS Code

Please try it yourself!

# Try with any issue/repo, I used gsm.qtl
# Get the appropriate branch
devtools::install_github("Gilead-BioStats/gsm.utils", "ai-playground")
gsm.utils::syn_gsm_standards()
gsm.utils::build_agent_prompt("gsm.qtl#78")

copy/paste the prompt into copilot, and proceed to write more detailed documentation on other issues as copilot gets stuff done

[Jeremy Wildfire (jwildfire)]

I like this in general! Tons of food for thought. I've added a couple of high-level comments below:

• This is probably too much for a first step. I'd lean towards scaling back and adding a small handful of things to the scaffold and then iterating
• We need a big readme/contributor guide update to go along with this that clearly explains how the humans set this up and what their respsonibilities are.
• Not entirely sure the R files are needed, but I might be wrong. I think we should generally prefer command line api calls (especially the gh github tool) and markdown templates so that we are language-agnostic as much as possible
• I'll paste parts of my current AGENT.md as a comment for your consideration. There's also a PR over in gsm.core with a draft from a few months back that might be useful reference: Gilead-BioStats/gsm.core#95

Next steps:

• Overall goal: Let's clearly align on scope of the first iteration of our gsm AI scaffold
  • The deliverable here is probably a "spec" or design document.
• Get feedback from other team members on the initial draft(s)

[Zelos Zhu (zdz2101)]

I like this in general! Tons of food for thought. I've added a couple of high-level comments below:

• This is probably too much for a first step. I'd lean towards scaling back and adding a small handful of things to the scaffold and then iterating
• We need a big readme/contributor guide update to go along with this that clearly explains how the humans set this up and what their respsonibilities are.
• Not entirely sure the R files are needed, but I might be wrong. I think we should generally prefer command line api calls (especially the gh github tool) and markdown templates so that we are language-agnostic as much as possible
• I'll paste parts of my current AGENT.md as a comment for your consideration. There's also a PR over in gsm.core with a draft from a few months back that might be useful reference: Gilead-BioStats/gsm.core#95

Next steps:

• Overall goal: Let's clearly align on scope of the first iteration of our gsm AI scaffold

  • The deliverable here is probably a "spec" or design document.

• Get feedback from other team members on the initial draft(s)

Yea I suspect we need some sort of "skill graph" of .mds that tells the story of the relationship between the packages and why they're related to each other, basically turning this into mds and potentially citing where each package is involved

[Zelos Zhu (zdz2101)]

Closing to keep repos clean and organized