Give any LLM the power to ask the right question.
normalpowers is a portable prompt bundle that turns ChatGPT, Gemini, or any flavor of Claude into a patient interviewer who asks the right questions, surfaces unexamined assumptions, and walks you through rigorous thinking on any knowledge-work problem.
Works on any LLM — paste the master prompt and go. Claude Code users can install in one command.
normalpowers is a fork of superpowers by Jesse Vincent. superpowers is the original — a Claude Code plugin that makes LLMs meaningfully smarter at software work by giving them a metacognitive layer: "before you act, check which skill applies."
Jesse's insight is the whole idea. We are deeply grateful the project exists.
What normalpowers changes:
- Target: every LLM, not just Claude Code. ChatGPT Custom GPTs, Gemini Gems, Claude Projects, Claude.ai free tier, Claude API, and anything else with a system-prompt slot.
- Scope: knowledge work, not software development. The skills we kept (
brainstorming,writing-plans,verification-before-completion,writing-skills) were already domain-neutral. We re-exampled them for product, strategy, research, writing, ops, sales, and more. - Stripped:
test-driven-development,executing-plans(replaced byfollowing-your-plan),using-git-worktrees,subagent-driven-development,requesting-code-review,dispatching-parallel-agents,finishing-a-development-branch, and all plugin/hook/tool-system machinery.
If you build software, use superpowers — it is tuned for that and sharper than anything we could offer. If you work on everything else, try normalpowers.
See docs/comparison-to-superpowers.md for a full side-by-side.
Eight skills that shape how your LLM approaches knowledge work:
| Skill | Use when… |
|---|---|
brainstorming |
You're figuring out WHAT to build, write, decide, or plan. Before drafting. |
writing-plans |
You have a clear target and need to break it into executable steps. |
following-your-plan |
A plan exists and you're executing it. |
systematic-problem-solving |
Something is broken, underperforming, or unexpected. You want to understand why before fixing. |
verification-before-completion |
You're about to claim "done" or "ready". Verify with evidence first. |
receiving-feedback |
You got critique and need to evaluate and respond. |
writing-skills |
You want to author a new skill for your own workflow. |
using-normalpowers |
The entry-point skill every conversation starts with. |
The LLM is instructed to check which skill applies before responding — even before asking clarifying questions. This is the whole trick.
Copy normalpowers.md from this repo. Paste it into whichever platform you use.
Install guide priority: Claude first (it's our internal stack), then ChatGPT and Gemini.
/plugin install normalpowers@fastxyzOr download normalpowers-v0.1.0.plugin for a local offline install:
/plugin install ./normalpowers-v0.1.0.pluginFor skill-compatible platforms, download normalpowers-v0.1.0.skill.
- Open claude.ai → Projects → New project.
- In the project's Custom instructions field, paste the full contents of
normalpowers.md. - (Optional, improves behavior) Download
normalpowers-v0.1.0.zip, extract it, and drag the extracted folder into the project's knowledge section. Or upload the individualskills/*/SKILL.mdfiles from GitHub if you prefer. - Start a chat in the project. Say "I want to plan a hackathon" — Claude should enter brainstorming mode, asking one question at a time, rather than jumping to a plan.
- Start a new conversation.
- Paste
normalpowers.mdas your first message. - Continue the conversation from there. The behavior persists for the session. Repeat the paste for each new session.
Pass normalpowers.md as the system parameter:
from anthropic import Anthropic
with open("normalpowers.md") as f:
system_prompt = f.read()
client = Anthropic()
response = client.messages.create(
model="claude-opus-4-7",
system=system_prompt,
max_tokens=4096,
messages=[{"role": "user", "content": "I want to plan a hackathon."}],
)For deeper behavior, concatenate selected skills/*/SKILL.md files after the master prompt.
- ChatGPT → Explore GPTs → Create.
- In the Configure tab's Instructions field, paste
normalpowers.md. - Under Knowledge, upload the eight
skills/*/SKILL.mdfiles so the GPT can retrieve them when the master prompt references a skill by name. - Save. Start a conversation.
If the GPT ignores the skills, try pasting the master prompt as the opening user message instead — Custom GPT Instructions are sometimes deprioritized for long prompts.
- Gemini → Gems → New Gem.
- In Instructions, paste
normalpowers.md. - Gems don't support file retrieval, so the individual skill files must be inlined. For full-fidelity behavior, paste the relevant skill file into the chat when the Gem announces "Using
<skill>".
Known issue: Gemini sometimes ignores long instruction blocks. If you see it skipping straight to answers without the skill-check preamble, shorten the master prompt or paste it as the first chat message.
Any LLM with a "system prompt" or equivalent (Mistral, Llama variants, local models): same pattern as Claude API. Paste normalpowers.md as the system message.
After installing on any platform, try this conversation:
You: I want to plan a launch for our new onboarding flow next month.
LLM (with normalpowers): Using
brainstormingto sharpen this launch plan.Before I suggest anything, a few questions — one at a time.
Who is the primary audience for the new onboarding flow — existing customers, new signups, or a specific segment?
Without normalpowers, most LLMs will produce a plausible-looking launch plan template in the first response. That's the failure mode normalpowers exists to prevent.
See examples/ for longer transcripts.
"Normal" thinking is hard. Asking the right question, sharpening an idea before producing anything, decomposing into small verifiable steps, checking evidence before declaring done — these are the habits that distinguish a careful senior operator from a frantic junior one. LLMs default to frantic junior. normalpowers makes them act like careful seniors.
Full thesis: docs/philosophy.md.
Please read CONTRIBUTING.md before opening a PR. Short version:
- Issue-first for non-trivial changes.
- One problem per PR, tested on a real LLM session.
- Domain-specific skills (sales-discovery, 1:1-prep, specific-industry-workflow) belong in your own bundle, not core. The
writing-skillsskill is the entry point for authoring them. - Platform integrations, rebrands, and tone sweeps without a behavioral problem get closed.
normalpowers deliberately mirrors superpowers' directory structure where we kept a skill unchanged in shape, so that git cherry-pick from obra/superpowers remains tractable. See docs/upstream-sync.md for the sync workflow.
MIT License. Jesse Vincent's 2025 copyright is retained verbatim in LICENSE. Pi Squared Inc. and normalpowers contributors hold 2026. Derivation is documented in NOTICE.
- Jesse Vincent and the team at Prime Radiant — for superpowers, and for the insight that makes this whole approach work.
- Pi Squared Inc. / FAST — for the room to build this and the internal dogfooding that shaped the knowledge-work examples.