Conference/ORCwAI/Pt2/1220 Harper Reed LLM Codegen
1220 My LLM Codegen Workflow at the Moment - Person/Harper Reed
- 12:30 this is definitely based on his recent blog posts
- 12:30 AI/Coding/v0/File/TICKET-spec.mdc - keystone of entire process is the spec
- if you don't have the spec it won't work
- "lately it's gotten so good I don't even read them"
- 12:31 feed spec into reasoning model, and have it generate the prompts
- 12:32 Greenfield Coding loop
- 12:32 "Context packing"
- mentions repo2txt
- one test at a time, one feature at a time
- do a new branch - use it like a video game checkpoint
- he mentions pre-commit Py/Lib CICD/CI on every commit
git reset --hard every four hours- 12:33 "Waterfall" is no longer evil
- LLM needs clear specs
- "the micro-cycle is waterfall"
- you can do it concurrently
- 12:35 no more jobs we can recognize
- 12:36 pro tips
- start plain
- scaffolds keep models honest
- short prompts beat essays
- 12:38 closing remarks
- 12:46 some one mentions this