Giving the Agent a Working Method: GrillMe & Superpowers

Advanced section. The setup guides gave your agent a brain (the university model) and hands (the terminal). This page adds a working method: two freely available skill collections that make the agent plan before it builds, and question you before it plans.

What this adds

Without it, a capable agent answers your request with code as fast as it can – including the code for the thing you did not actually mean. With it, the agent follows a discipline that experienced developers use on themselves:

Both work with the university model you already configured. Nothing new to install beyond the steps below, and both can be removed again with one command.

The method in one picture

Interview → spec → plan → build → check. First the agent questions you (GrillMe if you ask, brainstorming automatically with Superpowers), until the idea is sharp. Then a short design you approve in readable chunks. Then a plan written so that a stranger with no context could follow it. Only then code – built piece by piece, each piece reviewed against the plan.

The order is the point. Most wasted agent time is not bad code – it is good code for the wrong plan. These skills move the thinking to where it is cheap.

Install on Hermes

Prerequisite: the Hermes setup guide is done and Hermes is running. Open a new Hermes session and type (the agent will ask for confirmation before running anything – approve it):

Run this and show me the result:

hermes plugins install obra/superpowers --enable

This is the install command from the official Superpowers documentation. Then end the session and start a new one – a running session does not pick up a freshly installed plugin.

Now GrillMe, same session, one command:

Run this and show me the result:

hermes skills install official/software-development/grill-me

Plan B (manual): if the agent is unreachable, run the two commands yourself in a terminal. They are exactly the lines in the boxes above.

Install on OpenCode

Prerequisite: the OpenCode setup guide is done (OpenCode 2.0.4 or later – the version from the guide is far newer; check with opencode --version). The plugin arrives via Git, so first check that Git exists:

git --version

If that says “command not found”: install Git from git-scm.com (default settings are fine), then open a new terminal window.

Now tell OpenCode to install it. This sentence is the officially documented route – the agent fetches the instructions from the source and does the edit itself:

Fetch and follow instructions from
https://raw.githubusercontent.com/obra/superpowers/refs/heads/main/.opencode/INSTALL.md

OpenCode will read the instructions and edit your opencode.json – approve the edit when it asks. One line is added to the file in your home folder:

"plugins": ["superpowers@git+https://github.com/obra/superpowers.git"]

Quit OpenCode and start it again – the plugin loads at startup.

Now GrillMe. OpenCode has no skill registry, so the agent creates the file directly (approve the file creation):

Create the file ~/.config/opencode/skills/grilling/SKILL.md with the exact content of
https://raw.githubusercontent.com/mattpocock/skills/main/skills/productivity/grilling/SKILL.md
Do not change the content. Show me the file afterwards.

Plan B (manual): open opencode.json in your home folder and add the “plugins” line from above at the top level (with a comma after the neighbouring entry, none at the end). For GrillMe: download the file from the URL, create the folder .config/opencode/skills/grilling/ in your home folder, save the download as SKILL.md inside it.

First use

Two tests, in a new session (Hermes) or a restarted OpenCode:

1. Is it there? Type: Tell me about your superpowers. The agent should describe its skills instead of looking puzzled. (This check comes from the official Superpowers install docs.)
2. Does GrillMe bite? Give it a real idea of yours and the trigger phrase: I want to build a small website where students can browse a text corpus. Grill me on this plan. You should get numbered questions with a recommended answer each – answer, and the next round builds on what you said.

After that, everyday steering looks like this:

What you actually gain – honestly

• More questions at the start. The first ten minutes feel slower. That is the product: decisions move from “after the code exists” to “before”.
• A plan you can read and veto. Nothing is built until you have approved a description in plain language.
• Built-in self-checking. During execution the agent reviews its own work against the plan instead of declaring victory after the first version.
• The price: it is overkill for a two-line one-off task. For anything you would hate to do twice, it pays for itself in the first revision it prevents.

Safety: what are you installing?

Skills are instructions the agent will follow – see rule 3 on the safety page: outside content is treated as hostile until you have looked at it. So look at what you are adding here:

Both install through your confirmation gates – nothing runs or writes without your approval, and your safety settings stay exactly as they are. If you change your mind:

• Hermes: hermes plugins remove superpowers and hermes skills uninstall grill-me
• OpenCode: delete the “plugins” line from opencode.json, and the folder .config/opencode/skills/grilling/ – both gone on the next restart.

If something does not work