How I AI: My Weekly Codex Experiments

TL;DR

• 1

  Codex can build its own context window from your filesystem — Nate describes asking Codex to find files by what they’re about and roughly when they were made, then copy them into a neat working folder instead of relying on exact filenames.

• 2

  That workflow unlocks long, messy projects — with a clean local folder as context, he says Codex can reliably handle 30,000-50,000 words, plus complex document, spreadsheet, and coding tasks across multiple files.

• 3

  Claude didn’t hold up the same way in this exact setup — Nate says he tested the same workflow with Claude Code/Claude Co-Work and it “doesn’t work,” though he leaves open whether that’s model quality, long-running task reliability, or compute limits.

• 4

  Prompting has shifted from task assignment to collaborative scoping — instead of saying “go do this,” he now gives a set of meaningful questions, points to relevant files, and asks the model to help define the task shape before execution.

• 5

  Long-running agents are finally staying on track — especially with Claude 5.5 and refreshed Codex, he feels he can shift from messy back-and-forth exploration to “now go do it” without the model getting lost.

• 6

  Multi-threading is the real unlock — Nate says he can now run simultaneous drafting in local folders and even develop eight or nine prompts to execute sequentially, helped by Codex’s file handling and auto-review guardrails.