# Starter prompt: ChatGPT and Codex as planner, explainer and second worker

<!--
What this file is: two prompts. Part A goes into a normal ChatGPT conversation (planning, explaining
errors, reviewing configs). Part B is a brief for Codex CLI running on the server with a ChatGPT plan,
in its own folder. Edit the bracketed values first. Guide section: doc 2, "Setting this up with limited
experience". Template only.
-->

## Part A: paste into ChatGPT (planning and review, no server access)

---

I am building a self-hosted AI server and want you as my planner and reviewer. Another assistant
(Claude through Desktop Commander) runs the commands; you do not have access to the server.

My setup: [provider], [CPU], [RAM] GB, [GPU or no GPU], Ubuntu [version] LTS, domain [example.com] on
Cloudflare. Goal: Ollama for small local models, LiteLLM as the single model gateway, Qdrant for memory,
Cloudflare Tunnel plus Access for every web page, Guacamole for a browser desktop, systemd timers for jobs,
ntfy for alerts, and a plan-then-run prompt hub.

What I want from you:

1. A build plan in layer order with a one-line reason for each layer, and what I must decide before each.
2. When I paste an error (I will remove secrets first), explain it in plain words and give the most likely
   cause and a read-only check that confirms it.
3. When I paste a config file (secrets replaced by placeholders), review it for: ports bound to 0.0.0.0,
   missing version pins, missing health checks, secrets in the file, and anything that does not match
   the official docs as of today. Say which official page you are relying on.
4. Never ask me for passwords, keys, tokens or ntfy topic names.

---

## Part B: brief for Codex CLI on the server (its own folder)

Run it from a folder that belongs to Codex only, for example:

```bash
mkdir -p ~/agents/codex/2026-10-15-review && cd ~/agents/codex/2026-10-15-review
codex exec --sandbox read-only --skip-git-repo-check - < ~/agents/briefs/codex-review.md
```

Contents of `~/agents/briefs/codex-review.md`:

---

You are reviewing a self-hosted AI server that another assistant is building. Work read-only.

1. Inventory: OS release, kernel, CPU flags (avx2, avx512, amx), RAM, disks, listening ports
   (`ss -ltnp`), running containers (names, images, published ports), enabled systemd units and timers.
2. Flag any port listening on 0.0.0.0 or [::] other than SSH, any container image without a pinned tag
   or digest, any unit without Restart=, and any file under the home folder with mode wider than 600
   whose name suggests a secret. Report file NAMES only, never contents.
3. Compare what you find against the build plan in ~/setup-log.md and list gaps.
4. Write your report to ./RESULT.md in this folder. First line: "PASS" or "GAPS FOUND". No secrets,
   no IP addresses, no tokens in the report.

Do not modify anything outside this folder. Do not restart services. Do not install packages.

---