A knowledge kit for AI coding agents
Your agent reads PowerWorldHiveMind before it touches PowerWorld Simulator, so it knows the calls that report success and change nothing. Stop babysitting it, and stop being the documentation.
Built for Claude Code. Codex, Cursor, Windsurf and Copilot read its AGENTS.md, and Gemini CLI reads a copy, GEMINI.md: checked against their docs, not tested on each one. You need Windows, PowerWorld Simulator and its SimAuto add-on.
22 of 23 PowerWorld questions right with the kit, against 13 for the same model searching the web, marked blind.
The agent never reads the whole kit. It turns your question into search terms, searches every page at once, and reads only the page that matches.
A replay: real searches over the kit, run when this site was built. The search terms and the answer are written by hand.
The same model answered the same 23 PowerWorld questions twice: once with this kit on disk, once searching the open web. A grader who did not know which was which marked each answer right or wrong.
Closing a breaker opens that section.
About fifteen minutes, and no coding. Three downloads, a few clicks, and one sentence pasted into a chat box. You need PowerWorld Simulator on this Windows computer already.
Download it, open the file it saves, and sign in. You need a paid plan (Pro or above), because the free plan doesn't include the Code tab. Install Git too, clicking through with every default. Claude uses it to fetch the kit.
Run the installer, and on its first screen, tick "Add Python to PATH" before you click Install. The box is off by default, and missing it is the most common way this goes wrong.
In Claude, click Code at the top left. The Chat tab can't see your files. Code can, and it's where everything below happens.
Paste this into the message box and press Enter. It's a request, not a command: you're asking Claude to fetch the kit for you.
Install the PowerWorld knowledge base: git clone https://github.com/ChunSikPark/PowerWorldHiveMind ~/.claude/skills/powerworld-hivemind
When it's done, type this to load it:
/reload-plugins
Type this. It installs what the kit needs and checks that PowerWorld can be driven from this machine.
/powerworld-hivemind:powerworld-setup
If it says SimAuto isn't licensed, that's a PowerWorld licence question for whoever manages yours, not a mistake you made.
Use Claude Code in a terminal? Setup also offers a PWHM dashboard for your status line: the kit's version, the model, your usage limits and how full the context is. It asks first. If you already have a status line, it keeps yours and adds just the [PWHM#0.4.0] tag. The desktop app doesn't show status lines.
Start with this one. The right answer is "it silently writes no file". If it says "it saves the case", type /reload-plugins and ask again.
PowerWorldHiveMindWithout running any code, what happens if I call pw.esa.SaveCase("out.pwb")?Then point it at one of your own cases, in plain English:
Open my case at C:\path\to\case.pwb and summarize it.
Run an N-1 contingency analysis and show me the worst violations.
You won't have to check for updates. When a new version comes out, Claude tells you at the start of a session and offers to update it.
Stuck on a step? The getting-started guide shows what each failure looks like and how to fix it.
Through the Codex plugin marketplace.
codex plugin marketplace add ChunSikPark/PowerWorldHiveMind
Then open /plugins, install powerworld-hivemind, and start a new session.
Not yet tested on a released Codex build. If it fails, use the clone route in the other card: Codex reads AGENTS.md too.
Any agent that reads the files in a folder.
git clone https://github.com/ChunSikPark/PowerWorldHiveMind cd PowerWorldHiveMind pip install esapp TeamOverbyeWeather
Start your agent in that folder and tell it: Read AGENTS.md and run the preflight.
You talk to the agent in plain English. The agent reaches Simulator one of two ways.
The agent writes Python with esapp, which drives Simulator through SimAuto. It opens the case, writes changes, solves, and reads every write back, because PowerWorld reports success on writes that did nothing.
Fast, scriptable, runs in parallel. Every demo uses it. Needs the SimAuto add-on, which is licensed separately.
Turn on one setting and Simulator watches a folder. The agent writes an .aux script and drops it in. Simulator runs it and writes back the result CSVs and its own message log for that script: records read, control adjustments, every mismatch iteration, and whether the solve succeeded.
A SimAuto call hands back only an error string, so a silent failure stays silent. Here the agent reads the log line that explains it, and you keep a script you can read and re-run.
It's slower: a dialog has to stay open, so there's no batching, and it needs a recent Simulator build.
Setup, rules and a templateThe agent never reads the whole kit. It turns your question into search terms, searches every page at once, and reads only the page that matches. The searches below ran on the kit itself when this site was built.
Your words, and the PowerWorld names they probably map to. A page might use either, so the agent searches both.
A few lines that say whether this is the right page. If it isn't, the page's Connections say which one to open instead.
Why search first: on 16 test questions, searching the pages found the right one 10 times, against 8 for the routing table and 7 for starting at the index. The full search rules
A study is more than one command, so the work is split into six roles. Two are live: schema-librarian and case-auditor. The other four are planned and not built yet, and three of them will answer with a page instead of a wall of text.
planned network-visualizer will answer with
Your design and up to three challengers on the map, each measured on the full N-1. Pick one and the map redraws with its overloads painted on.
planned run-supervisor will answer with
A long N-1 or time step, live: a status badge, progress, a run clock and a timeline. It will alert your phone if the run stalls.
planned plan-pictures will answer with
Before you approve anything, the plan will be drawn as a swimlane from the plan the engine wrote from your settings, safety checks included. If the settings change after that, it won't draw a stale plan.
Every role, live or planned, is designed to work from a plain request, ask before it changes anything, and never save over your case.
Not built yet: this is how it will answer.
Most tooling answers "which branches are overloaded" and stops. The useful work comes after that readout.
In the remediation demo, two of five plausible fixes made the system worse, including the one an engineer would pick first. You only find that out by measuring.
The same 23 questions split by kind, and the case-auditor agent's own blind-judged test.
Web search finds what a command does. It rarely finds that the call reports success and does nothing: on those 16 questions the web got 7 and the kit got 15.
The kit also needed 3.2x fewer searches per right answer.
The case-auditor agent and a plain model got the same audit findings for 8 cases, 3 runs each. This test ran on a build after v0.4.0, with Sonnet. A blind judge found the agent described the case correctly in 21 of 24 runs, the plain model in 10. The plain model hardly ever invents anything: on one combined check (no invented cost data, only the three severity labels, no pasted raw findings) it passed 50 of 51, the agent 51 of 51. The difference is in the call: what actually stops the study.
For exact argument syntax, open PowerWorld's manual instead. The kit deliberately does not publish it, so it sometimes names the right command without a signature.
Every question, both verdicts, and what each setup spentAnything you work out and leave in the chat is gone when the session ends. Say so and the agent writes it down.
write that up as a page
document this gotcha before I forget it
Or run /powerworld-hivemind:kb-page. Not on the plugin? Tell your agent to read skills/knowledge-base-page/SKILL.md. It is one markdown file.
Your copy stays yours. A clone does not sync back, so nothing you write is published unless you open a pull request.