Your AI, but it actually knows how to investigate, not just chat.
You type: "Track the latest papers on brain-to-text neural decoding"
Instead of a vague reply, you get a structured report: 10 key papers from the last 3 months, each broken down into what problem it solves → how it solves it → what the results actually mean → what the catch is. With links. With a note on which papers contradict each other. With a "if you only read one, read this" recommendation.
You type: "Is that viral story about company X actually true?"
Instead of summarizing tweets, your AI cross-references court records, SEC filings, and independent media. It labels each claim: 🟢 verified / 🟡 likely / 🟠 disputed / 🔴 unverified. It tells you which sources it trusts and why. And it admits what it couldn't find, with suggestions on where to look next.
That's what this repo does. Not better prompts — better ways of thinking, baked into instructions your AI follows step by step.
Six "skills" — each one is a detailed workflow (200-500 lines) that turns your AI from a chatbot into an investigator. They work on Claude Code out of the box. They also work on ChatGPT, Cursor, OpenClaw, and anywhere else you can paste a system prompt.
| Skill | Try this (EN) | 或者说 (中文) | What you get |
|---|---|---|---|
| source-sleuth | "investigate: is [claim] true?" / "fact-check [story]" | "溯源：XX事件是真的吗" | A confidence-rated report tracing every claim back to its original source — with a "here's what I couldn't verify, go check these places yourself" section |
| research-tracker | "track papers on [field]" / "latest research on [topic]" | "跟踪 neural decoding 最新论文" | ~10 recent papers, each explained in plain language with a non-academic analogy for the method, plus a map of how the papers relate |
| domain-survey | "survey [field]" / "what is [domain]?" / "field guide to [topic]" | "帮我入门 量子计算" | A structured field guide: what it is → why it exists → how it works → what people fight about → a glossary where every abbreviation gets spelled out |
| news-summary | "daily news" / "tech news today" / "news digest" | "今天有什么新闻" | A categorized daily digest (tech / finance / world / China), each story with source links, automatically flagging breaking news |
| strategist-analyst | "analyze the strategy behind [event]" / "who benefits from [situation]" | "分析XX事件的博弈" | A dual-mode analysis (analyst: who holds what cards / detective: what chain of events led here) based on real strategy frameworks |
| book-deconstructor | "deconstruct [book]" / "break down [book]'s arguments" | "拆解《书名》" | The book's skeleton: core argument → evidence chain → hidden assumptions → what the author conveniently left out |
You want to understand something new
│
▼
domain-survey ← draws the map
(terms, subfields, debates)
│
┌────────┼────────┐
▼ ▼ ▼
research- source- strategist-
tracker sleuth analyst
(stay (trace (analyze
current) truth) the game)
They share what they learn. Domain-survey's glossary feeds research-tracker's search terms. Source-sleuth's verification rules echo through every other skill.
git clone https://github.com/KIRRAWA/cogniforge.git
cd cogniforge
./scripts/install.sh --allRestart Claude Code. That's it.
Pick and choose:
./scripts/install.sh source-sleuth research-tracker # just the investigation tools
./scripts/install.sh strategist-analyst book-deconstructor # just the thinking toolsThese are plain text files. The only Claude Code-specific part is 3 lines of YAML at the top. Strip that, and the rest is a universal system prompt.
Same file works on:
- Claude API — drop the body into
systemparameter - OpenAI Custom GPTs — paste into Instructions box
- Cursor / Windsurf — copy to
.cursorrulesor project rules - OpenClaw — native support, just copy the folder
- Any LLM web app — paste as the first message in a conversation
- LangChain / agent frameworks — use as system message
See full migration guide with code examples →
(coming soon — for now, the pattern is: remove the --- block at the top, use the rest as system prompt)
Most "AI prompt collections" give you sentences like "Act as a teacher."
These skills give you:
- Step-by-step workflows — Phase 1 → Phase 2 → Phase 3, with parallel branches where things can run simultaneously
- Source tiering — an S-through-F ranking of how much to trust each type of source (court records = S, Reddit comments = F). Built into every skill that touches external information.
- Predefined output structures — not "write whatever," but specific sections with word limits
- Self-check checklists — each skill has a "before you deliver, verify these 5 things" list
The closest thing on GitHub is awesome-chatgpt-prompts (166k ⭐). That's a collection of one-sentence role-play prompts. This is a collection of full cognitive workflows. Different category.
cognitive + forge — a furnace for turning raw information into structured knowledge.
Also: platform-agnostic by design. These workflows aren't married to Claude, GPT, or any single model. The thinking is the asset; the platform is just where it runs.
See CONTRIBUTING.md. The short version: a good skill has a clear job, step-by-step phases, source quality rules, an output template, and a self-check list. Use skills/domain-survey/SKILL.md as a reference.
CC BY-SA 4.0 © 2026 Nuofan Yang — use it, modify it, sell it. Just credit the source and keep derivative works open under the same license.