Agents' review of our project :)

Claude Code, 4/6/26

Prompt: "Hi, claude! can you take a look at my repo and give me your take?"

⏺ This is a solid, well-conceived project. Here's my take:                                                                                  
                                                                                                                                            
  <strong>What's working well: </strong>                                                                                                                     
                                                                                                                                            
  - <strong>Architecture is clean</strong>. Single source of truth in AGENTS.md, agent-specific files as thin redirects, skills as on-demand loads. That's the right call for avoiding content drift.                                                                                                
  - <strong>Task-oriented naming</strong> (writing-automated-tests, not tests) is genuinely better for agent task-matching.                                  
  - <strong>Priority-ordered options</strong> in skill files is a subtle but important design choice -- agents that default to the wrong test type is a real failure mode, and the ordering nudges them correctly.                                                                                     
  - <strong>Eval framework</strong> is the most impressive part. Most prompt engineering projects skip validation entirely; this one has behavioral + static checks that can prove whether guidance actually changes model output.                                                                     
                  
<strong>  Things I'd think about: </strong>

  - how-to-write-documentation feels thin -- if it's a stub, it might be worth either fleshing it out or clearly marking it as status: draft
   so agents don't pull it and get incomplete guidance.
  - The redirect files (CLAUDE.md, GEMINI.md) could become stale if new tools emerge. You might want a note in CONTRIBUTING.md about keeping SETUP.md updated when adding new agent tool support.                                                                                     
  - Eval coverage looks solid for coding-standards but the writing-automated-tests structure exists -- are evals actually filled in for that
   one? If not, there's a gap between what's validated and what's shipped.                                                                                                                            
                  
  Overall: this is one of the more thoughtful "AI context for a specific platform" projects I've seen. The eval-driven approach in          
  particular is ahead of where most teams are.