Write a prompt contract the model can quote back: inputs, output shape, tool names, error policy, one example. Check it into the repo. Treat changes as breaking until evals pass.

Codex is GA with SDK and Slack integration. If your org lives in GitHub and Slack, wire those first. The quickest wins are routing and reviews. citeturn0search2turn0search4

Default settings for sane automation: - reasoning_effort: low - verbosity: brief - scope: 1-3 files - budget: N tokens or 60 s - success: tests green Escalate only after two failures. You pay for uncertainty. Shrink it, then spend where it matters.

CLI autonomy is a dial, not a switch: suggest mode to preview auto-edit to confirm changes full-auto when tests guard you Pick the lowest autonomy that still ships.

From papers to PR: inputs: goal, repo scope, budget tools: fetch_doi, extract, run_tests output: patch with inline references If it cannot cite, it cannot merge.

One run, one model. The model that plans update_linear_task(...) is the one that executes it. Swaps only happen on the next turn, or when you use the router (gpt-5-chat-latest) or ChatGPT. Use reasoning_effort to trade speed vs depth. citeturn6search0turn5search0

Steerability shines when you budget it. Game loop: - tool budget - frame budget - test budget Start minimal reasoning for hot paths; bump to medium/high for tricky AI. Log preamble and every tool call, then diff before merge.

Software meets silicon. Prep your stack now: - stream by default - batch calls - keep context tight - reasoning_effort low on hot paths, raise on hard cases - measure tail latency When compute improves, you’ll be compounding.

Two dials: verbosity = how much to say reasoning_effort = how hard to think Lint-level edits: brief + low. Gnarly bugs: balanced + high. Log both per task and tune.

Codex-era lesson: say less, show more. For GPT-5, try: <goal> speed up the loop </goal> <files> parser.ts </files> <tests> keep API stable </tests> Start with --suggest. Promote to --auto-edit when green.

Choose the right brain in Codex: default GPT‑5 switch to gpt‑5‑codex for coding In‑session: /model CLI: codex --model gpt-5-codex. citeturn1view0

Encode taste once. Create ~/.codex/instructions.md with naming, comments, tests, and review rules. Codex follows your house style without prompt spam. Then tune verbosity and reasoning_effort per task.

Headless box? Do this: 1) Run codex login on a machine with a browser 2) Copy $CODEX_HOME/auth.json to the server (defaults to ~/.codex) 3) Use the CLI as normal Rotate creds? Replace the file. Clean and done.

New to Codex? Sign in with ChatGPT. Plus, Pro, or Team unlocks GPT-5 at no extra cost. fileciteturn2file0turn2file1 Headless or CI? Use an API key login. Upgrading to 0.20+? Delete ~/.codex/auth.json, then codex login.

Turn tests into tools. Describe them in plain text, then let GPT‑5 call them during edits. Keep tools deterministic and return JSON. Ship when green.

Context diet: Include only files in the loop, a short goal, and an Invariants block. Exclude vendor, build, and large logs. Less noise = faster, safer edits.

Better edits, tighter loop: 1) Ask GPT-5 for a plan 2) Get a diff preview 3) Run tests as a tool 4) Apply or abort Treat --suggest as the default and escalate only when green. fileciteturn0file1turn0file2

Two knobs to drive Codex: reasoning_effort: 0 for speed, higher for hard problems verbosity: low for brevity, high to teach Set them per task, not globally.

For migrations, go mechanical. Ask GPT-5 for a rewrite rule first. Apply to one file. Run tests and linters as tools. Batch only after the rule survives edge cases.

Use the right GPT‑5 variant: • gpt‑5 for critical reviews • gpt‑5‑mini for iteration loops • gpt‑5‑nano for small tasks Track tokens to control cost on long runs. fileciteturn0file1turn0file2