Why AI Agents Fail at Cross-Repo Code Design - And How to Fix It
AI agents write code fast but design badly across many repos. The fix isn't a smarter model. It's giving the agent a way to query the codebase instead of walking it.
> Complexity is often just an explanation problem.
# the SEMS engineering algorithm
def ship(code):
while not simple(code):
code = refactor(code) # delete beats add
return "🎉 ship it"
$ ship(production)
> complexity detected: 9001
> applying SEMS…
> ✓ simple. shipped.AI agents write code fast but design badly across many repos. The fix isn't a smarter model. It's giving the agent a way to query the codebase instead of walking it.
The /goal command sets a completion condition and Claude keeps working until a separate evaluator says it's done. Here's how to write goals that actually finish.
Launchpad is a Claude Code plugin that turns feature work into four guided slash commands - brain, architect, implement, test - with state that survives /clear.
Stop paying API rates to call your own assistant. claude -p turns a Claude Code subscription into a scriptable agent that already knows your skills.
A short, honest playbook for cutting Claude Code costs without switching to a worse tool.
BMad Builder generates the agents, skills, and workflows your team would otherwise hand-write. Here is the short version.
47,000+ agent skills across 6,300+ repos. Almost none tested beyond a vibe check. Here's a 4-layer testing architecture that fixes that.
How we built Automan - a static, YAML-driven internal portal entirely with Claude Code, no servers, no frameworks.
An internal framework for packaging team domain knowledge, workflows, and automation into reusable Claude modules - shareable, versioned, and improvable.
Stop arguing about quotes and line breaks. Black eliminates formatting debates entirely and gives back hours each week.