KM4BI KM4BI

How I use AI: my weekly Codex experiments

Nate B Jones explains a Codex workflow built around clean local folders, richer context assembly, and collaborative task shaping before agentic execution.

Nate B Jones is starting a weekly format about how he actually uses AI. This episode focuses on Codex and on a practical shift: agentic work becomes more useful when the local context is assembled into a clean working folder.

Context as a working folder

His workflow begins by asking Codex to inspect the file system and find documents described in natural language. He does not need to provide exact filenames. He can describe what the material is about, when he created it, or what kind of content it contains, and Codex gathers the relevant files into a dedicated folder.

He then starts a fresh chat focused on that folder. The model gets a cleaner context window, plus any detailed instructions or transcripts he adds as reference files.

Better suited to long work

This setup lets him handle 30,000- to 50,000-word document projects, spreadsheets, code, and structured prompt sequences. Nate connects this to Codex’s roots: it is comfortable reading files in a repo-like environment and reasoning about how they fit together.

In his current experience, the same pattern has not worked as well with Claude Code or Claude Code Work. He does not overstate the cause; it may be compute constraints, model differences, or the current state of the tools.

Prompting becomes collaborative scoping

The second shift is in prompting. Instead of immediately handing the model a task and a success criterion, he now starts with meaningful questions around the standards the work should meet. The model helps define the shape of the task before moving into execution.

Nate describes this as a more genuinely collaborative stage, especially with GPT-5.5 and the refreshed Codex: the model can help frame the work, then switch gears and carry it out without losing the thread.

Why it matters

The point is not vendor loyalty. Nate says he is not picking a side; he is following whatever makes him more effective. The important signal is that AI agents become more valuable when they can organize context, stay on task for longer runs, and operate with safeguards such as auto-review.

That combination enables more parallel work: drafting in multiple directions, incubating several ideas, preparing prompt series, and executing sequential tasks inside a well-structured local workspace.

Source