Every failed generation, every "that's not quite right," every wasted API call traces back to the same problem: you gave the AI garbage input and expected magic.
↓ Here's what we figured outThe problem isn't AI capability. Claude can write code, analyze systems, generate tests. The models are good enough.
The problem is you're skipping the hard part: figuring out exactly what you want.
A vague prompt produces vague output. A well-structured ticket with clear acceptance criteria? That produces code you can actually ship.
Every time an AI works on your code, it wastes tokens exploring. What if it already knew your patterns, your gotchas, your architecture? That's .llm-docs/—documentation optimized for LLMs, not us silly humans.
Before any code gets written, Claude helps you think through what you actually need. It asks the questions you'd skip. It finds the edge cases. It creates a ticket worth executing.
Point it at your backlog. It reads the ticket, writes the code, runs your tests, opens a PR. No babysitting. No "can you try again." It just works through the queue.
Because the input was right, the output is right.
Parses your GitHub Actions workflows and runs the same checks before opening a PR. If it fails locally, it fixes it—or doesn't waste your time with a broken PR.
Before opening a PR, it reviews each acceptance criterion and confirms it was actually met. No more "done" that isn't done.
Blocked by a PR that's still in review? It branches from that PR and keeps going. No waiting for merges.
When something breaks, it commits the work-in-progress with full context. You can see exactly what was done and resume from there.
Runs all night. Merges approved PRs automatically, pulls latest main, moves to the next ticket. Wake up to a cleared backlog.
Accumulates context from completed work and updates your .llm-docs/ after each ticket. The more it does, the better it gets.
The bottleneck was never AI capability. It's always been clarity of intent.
Thinking and doing are different modes. Trying to do both at once is why one-shotting fails.
Automation without structure is just faster chaos. Process isn't overhead. It's leverage.
I kept blaming the AI.
"It doesn't understand the codebase." "It missed the edge case." "It's not quite right."
Then I looked at my tickets. Really looked. And realized: I'd never actually said what I wanted. I'd been writing for developers who'd fill in the gaps, ask questions, course-correct as they went. AI doesn't do that. It follows instructions.
After twenty years of shipping software, I finally saw the job split in two:
ticket-worker is built for this split. You think. You evaluate. AI executes.
— Robert Beene
Select a ticket. Press enter. Review the PR when it's ready.