Back to Feed
How I AI: My Weekly Codex Experiments
The Signal
Nate, a frequent commentator on AI utility, argues that his productivity has surged after shifting from classic prompt engineering to a collaborative workflow where he shapes task parameters with agents before execution. He contends that his chosen tool, Codeex, enables this throughput by turning his local filesystem into an organized, clean context window for complex projects.
The Case
- Nate uses Codeex as a local-file assembler, tasking it to search his hard drive by natural-language description, copy relevant source files into a dedicated working folder, and execute tasks from that clean state.
- This workflow supposedly enables him to manage 30,000 to 50,000-word documents, dense spreadsheet work, and simultaneous coding projects, a feat he claims does not translate to Claude Code or the company's other 'co-work' offerings.
- He reports that his prompting style moved from structure-heavy engineering toward a 'messy' back-and-forth phase, where he queries agents about task shape and performance standards prior to committing to execution.
- By using these specific folder-bound workflows and an auto-review system that acts as a guardrail, Nate claims he can now 'multi-thread' eight or nine distinct prompts simultaneously to incubate multiple ideas.
- He remains explicitly uncertain whether these gains stem from recent Codeex updates, idiosyncratic differences between model versions '4.7' and '5.5,' or wider compute shortages reported across the field.
The 1 Minute Signal Take
The video is worth a watch if you are struggling with context windows and long-form AI throughput, as Nate describes a concrete, mechanical approach to file management that feels more actionable than generic prompting advice. Skip it if you are looking for definitive benchmarks, as he admits his findings are based on personal workflow observation and explicitly avoids pinning the performance changes on a specific model or version.
Time saved:
Back to Feed