Codex users share /goal audits, mobile delegation, and Raspberry Pi workflows

• OpenAI's Codex repo is public, and reach_vb's post framed a simple use case that kept showing up in the evidence: point Codex at its own codebase, then ask how features work or how the team itself uses the tool.
• /goal became the center of user workflows, with daniel_mac8's audit prompt turning it into a repo review pass, while reach_vb's Colab run showed it driving a browser session through a full training task.
• The most reusable prompt pattern came from reach_vb's 30-day memory prompt, which tells Codex to mine sessions, Memories, Chronicle, and existing assets before creating the smallest useful skill, subagent, or automation.
• Mobile use is shifting how people describe the product, with mattshumer_ on Codex Mobile talking about less micromanagement, while a translated mobile-control repost and TheRealAdamG's repost claimed iPhone task handoff still worked when the Mac was closed.
• Computer use is spilling far outside coding, from reach_vb's Raspberry Pi setup and thsottiaux canceling subscriptions to jxnlco's Cloudflare build thread, which had Codex creating a D1 database and editing Workers from Safari.

You can browse the source, skim the v0.133.0 release notes, and inspect community add-ons like agentmemory and openclaw/agent-skills. The weird bits arrived fast: gdb pushed Appshots, mattlam_ found a realtime_conversation flag and wired Codex into voice over WebRTC, and petergostev's post used remote computer use to buy and launch Fallout on a Mac.

The cleanest Codex workflow in the evidence is not "build X," it is "inspect what already exists, then turn the result into queued work."

According to daniel_mac8's prompt, /goal can run a comprehensive project review that asks for areas to improve, simplify, and align with the original plan. In the follow-up, that audit output becomes Linear issues via a Symphony server connection, so the review feeds directly into project management.

That same pattern shows up in lighter form in dkundel's tip, where /goal can be enabled mid-conversation instead of only at prompt start. The recurring idea across these posts is that users are treating Codex less like a turn-by-turn pair programmer and more like a long-running reviewer that can hand work off to another system.

The strongest prompt engineering thread here is about using Codex history as training data for future Codex behavior.

In reach_vb's expanded prompt, the source order is explicit:

1. Recent Codex sessions and task summaries
2. Codex Memories and rollout summaries
3. Chronicle, if enabled
4. Existing skills, custom agents, and automations

The output schema is explicit too. Codex should shortlist repeated workflows, attach supporting evidence and dates, estimate confidence, choose among skill, subagent, automation, extend-existing, or skip, then create only the high-confidence missing items reach_vb's prompt.

That is a more opinionated version of the shorter prompts in reach_vb's tip and the copy-paste template, which framed skills as reusable workflows and subagents as bounded delegated jobs like CI failures, PR reviews, changelogs, release prep, debugging, and test triage. dkundel's reply added a weekly refresh loop, and daniel_mac8's video post said the pattern produced eight recommended skills when paired with persistent memory files in Obsidian.

Alem Tuzlak, a TanStack contributor, supplied the most concrete downstream example in AlemTuzlak's /doc-writing skill, where a custom skill scans docs, finds insertion points for a new feature, writes for different user personas, and keeps documentation linked together as the codebase changes.

The mobile posts are less about screens and more about delegation style.

In mattshumer_'s post, the claimed behavior change is psychological: stepping away from the laptop leads to bigger prompts and less micromanaging. jxnlco's joke about no-laptop signs makes the same point from the opposite direction, treating the phone as a loophole for keeping an agent in the loop.

The product changes underneath that behavior were already surfacing two days earlier. The v0.133.0 release notes, linked in LLMpsycho's changelog post, listed goals on by default, foreground codex remote-control, permission profile APIs, plugin discovery, and extension lifecycle hooks. the reposted CLI instruction reduced remote access to a single codex remote-control command, which helps explain why so many examples in this evidence set jump between phone, desktop, and browser.

Users were already testing the edges. a repost from sengpt and TheRealAdamG's repost both claimed that iPhone task dispatch kept working even with the Mac closed, while swyx's complaint said the same update introduced too many approval pauses despite full-access mode.

Once users had /goal plus remote control, the use cases got weird in a hurry.

The most technical demo came from reach_vb's Colab run. Codex reportedly controlled Chrome in a signed-in Colab session, handled runtime and editing quirks, trained a roughly 10.65 million parameter transformer on a free T4 in about 19 minutes, then used subagents to audit the result.

Other posts pushed the same interface into more ordinary life-admin and ops work:

• thsottiaux used computer use to cancel subscriptions across different services.
• jxnlco watched Codex create Cloudflare Workers from the dashboard.
• jxnlco's follow-up said Codex set up D1, ran CREATE TABLE commands, added bindings, and edited Worker code in Safari.
• TheRealAdamG reposting JustinBleuel showed Codex driving the iPhone simulator to bug-bash a feature it had just built.
• petergostev said remote computer use bought, downloaded, and launched Fallout on a Mac.

The boundary line is now pretty thin. Browser automation, QA, deployment, dashboard ops, and consumer checkout all show up in the same evidence stream.

Several of the most believable hands-on reports were not benchmark claims at all, they were maintenance jobs on neglected hardware and software.

In reach_vb's post, Codex was setting up a Raspberry Pi and preparing Wi-Fi access; the follow-up said it successfully flashed the card and finished setup except for the Wi-Fi password. That is exactly the sort of chore where remote shell steps, config edits, and one annoying secret tend to dominate the work.

[Src:9|cramforce's thread] is the software equivalent. After a hand-built home automation stack bit-rotted around old Node 14 images, Codex reportedly modernized the whole system in three prompts and got deployment working again. The interesting detail is not the speed claim by itself, it is the target: a personal stack with Kubernetes, Docker, Helm, Next.js, InfluxDB, and self-hosted GitHub runners, which is the kind of legacy mess many agent demos quietly avoid.

The last cluster of evidence points to Codex becoming a broader interface layer, not just a coding terminal.

In mattlam_'s thread, a forked Codex app-server exposed a realtime_conversation feature flag and connected voice interaction over WebRTC to Codex agent turns. The claimed flow was: launch the app-server with realtime, connect a client over voice, let the realtime system decide when coding is needed, then hand those events to Codex for agent execution.

Appshots pull in context from the opposite direction. gdb's post introduced the feature, while a repost describing the shortcut said pressing both Command keys captures context from the active app and sends it into Codex. The public repo and changelog suggest the plumbing is increasingly extensible too, with the Codex source tree open for inspection and v0.133.0 adding plugin visibility and richer extension hooks.

That combination, open source client code, app-context capture, remote control, and experimental realtime wiring, is Christmas come early for coding agent nerds.