Why this note exists: For two years GitHub Copilot was the background hum in my editor. The shift to ChatGPT Pro + Codex turned that hum into an actual collaborator—one that reads context docs, syncs with GitHub, and can keep a roadmap moving when I'm away from my desk.
Why I moved beyond Copilot
Copilot is brilliant at inline autocompletion, but it rarely absorbed my broader intent. Codex, plugged into ChatGPT Pro, brings a deliberate planning layer. It reads the repo, cross-references series guidelines, and suggests tests before I even ask. The biggest gain is that architectural conversations, code edits, and documentation updates now happen in the same thread.
Direct GitHub sync keeps the agent accountable
The Codex CLI's tight GitHub integration eliminates the export-and-paste dance that slowed Copilot handoffs. I can pair an instruction with a target branch, have the agent open pull requests, and review the diff like I would a teammate's work. It also means I can ship from devices that never ran my full dev stack.
- Traceability: Every automated change points back to a conversation, so I can audit reasoning when regressions appear.
- Repo awareness: Codex respects CODEOWNERS, lint configs, and changelog formats without nudging.
- Faster reviews: Suggested commits arrive as cohesive stories, not a scatter of inline edits.
Learning what “vibe coding�really means
The first week I kept hearing people say Codex enables vibe coding—shaping software through conversation rather than specs. In practice, it's structured improvisation: I sketch goal posts, let the agent propose an approach, and then we iterate on the diff. Existing tools let me do this piecemeal; Codex makes it sustainable because it remembers house style, naming conventions, and the jobs-to-be-done for each component.
I still write the guardrails, but Codex keeps nudging the work back toward the “vibe�we documented—predictable, testable, easy to explain to collaborators.
AGENTS.md is the operations manual
The secret to stability is AGENTS.md. It's a living handbook that tells Codex how to behave: escalation rules, coding standards, testing philosophy, even how to talk about stakeholder trade-offs. Because the file lives in the repo, every session inherits the same ground truth. Updating AGENTS.md is the equivalent of running a company-wide training—one edit, immediate behaviour shift.
Right now my AGENTS.md covers:
- Language preferences for docs and commit messages.
- Required checks before touching regulated datasets.
- Escalation paths: what to do when permissions or tooling get in the way.
Copilot never really had a place to store this playbook. Codex reads it at session start, so I can trust long-lived conventions.
Codex logic vs. other assistants
Most co-pilots optimise for predictive text. Codex optimises for goal completion. Under the hood, it runs a deliberate loop—plan, execute, verify—rather than streaming guesses. That means it can decline sloppy instructions, ask clarifying questions, or decide to open an ADR instead of spamming edits. The win for me is fewer silent failures: either the agent ships something testable, or it tells me what blocked it.
Shipping updates from my phone
I can start a voice note on the Tube, ask Codex to draft a feature, and by the time I unlock my laptop the pull request is ready for review. The mobile ChatGPT app pipes instructions into the same workspace, so ideas don’t die in my notes app. This is especially useful for documentation: I dictate structure, Codex drafts the Markdown, and a quick review on desktop wraps the loop.
How TODO.md keeps the agent honest
The other backbone file is TODO.md. Codex treats it as the single source of truth for near-term tasks. Every time I log an observation, the agent updates TODO.md, reorders priorities, and cross-links to the relevant files. When I resume a session, Codex checks the list first before proposing new work, so nothing falls through the cracks.
- Context digestion: TODO items store investigation notes, so Codex doesn’t re-run the same discovery.
- Status signals: Completed items move to a changelog section automatically, which makes weekly summaries trivial.
- Team handover: Collaborators can skim TODO.md and pick up where the agent left off.
What sticks after the switch
Swapping Copilot for ChatGPT Pro + Codex wasn’t about better autocomplete—it was about aligning the agent with my workflows. With AGENTS.md and TODO.md anchoring behaviour, vibe coding becomes a reliable operating mode instead of a novelty. The next experiment is extending this setup to teaching assistants so classroom tooling benefits from the same guardrails.
If you're curious about replicating the stack, start by drafting an honest AGENTS.md, migrate your active TODO.md, and let Codex handle the first low-risk pull request. The difference shows up fastest in how calm your backlog feels.