This 5-step release workflow starts with don't change the code

On X, Morlex broke Lauren Tan's feature shipping workflow into five stages and turned it into a Grok Bot configuration. The source material describes Lauren as a SpaceXAI engineer; the workflow has been packaged into the pstack plugin, with 54 skills, under the MIT license. In Grok Bot, first run /add-plugin pstack, then use /setup-pstack to choose models. Morlex's division of labor is: grok-4.7 writes code, opus 5.5 makes judgments and handles the hardest changes. Go deep first, then let go—don't have the agent run on its own right away. The third step is to paste that block of instructions into the Bot.

The five-stage workflow itself isn't complicated:

1. Understand: first /how, then /why, and explicitly say not to change any code yet.
2. Design: /architect, /arena, /interrogate; prototype before planning.
3. Build: /poteto-mode; reproduce first, pin down behavior, and profile.
4. Verify: /verify; require video as evidence.
5. Ship: watch it run, /loop, get it merge-ready, and keep a decision log.

What's really worth copying is the constraints, not the command names. Don't write code before understanding is complete; don't write a plan before the prototype exists; each stage must be able to prove its work and be able to stop the line; stop after two failures and have the agent explain why. Morlex also recommends that before any loop runs automatically, do it manually once so you know what good looks like; give the agent the same tools; read the transcript and turn repeated failures into checklist items.

Don't turn prompts into rituals either. Giving the goal and completion criteria is enough. Examples from the source material: /poteto-mode add a --json flag. text output stays byte-identical. verify both forms. Another is that startup takes 1.8s on a fixture; trace and fix the measured cause, and provide a before/after comparison. You can also use principle names instead of long explanations, such as prove it works: run the real import flow and show the write records; subtract before you add: delete deprecated adapters first, then design what's left.

There's a more practical question in the comments. DannyCarpenter asks: if it doesn't tell us, how do we know which model it's actually using? In the screenshot, Grok Bot replies: it cannot say which model is running underneath or being switched to; it can only do research, writing, coding, calendar/email management, and connectors through cloud agents, and distribute work to other bots. Another commenter says token consumption is too fast and doesn't recommend it.

The flowchart can be saved, and the configuration can be copied outright. The model black box and costs won't disappear just because the workflow looks nice. Before using AI as a team, run it manually once, know what good looks like, and then decide which parts can be delegated. Source: Morlex's original tweet https://x.com/0xMorlex/status/2117247463119532278 , pstack docs http://github.com/cursor/plugins/tree/main/pstack .