Fractal Framework is an experimental wizard that helps anyone turn a real local problem into a source-cited advocacy website without needing to be a researcher, a designer, or a developer.
You simply give it a place, and it conducts the research, writes the site, and keeps source citations beside the factual claims it generates so you can inspect them yourself.
A generated site about Hualapai Valley groundwater depletion, with inline citations linked to source documents.
- A student who's watching something go wrong in your community — whether it's the water table, an aquifer, or a reservoir dropping, the school being underfunded, a local trail being paved over, or another "activism vacuum" — and wants to do something about it but doesn't know exactly where to start.
- An educator looking for a project that gives students a real-world deliverable they can show their family and put on the web.
- A volunteer or adult advocate who cares a lot about a local issue but simply doesn't have the time, technical skill, or design palette to single-handedly design a website that's ready to take on the issue.
- Anyone who's simply tired of low-quality AI-generated content flooding the internet on topics they care about, and wants websites designed with actual editorial discipline and a visible source trail.
You type a location into the wizard — say, "Kingman, Arizona" — or just your hometown.
The wizard then:
- Identifies the local issue worth working on. A research agent loaded with the framework's methodology looks at your location, conducts in-depth research, and proposes the most urgent neglected issue ("activism vacuum") affecting that place right now. It explains its reasoning and shows you which sources it drew from.
- Lets you review and edit before locking anything in. The proposal is not treated as final. You can question the reasoning, redirect the issue, sharpen the framing, ask for a different direction, or start over. The campaign does not move forward until you approve the direction.
- Runs deep research against trusted sources to produce the campaign content — a hero pitch, key facts, an about essay, what's at stake if nothing changes, ways for readers to help, and a full bibliography.
- Carries citations into the site. Hash citations from the research output stay beside the claims they support. When a citation resolves to a source URL, its chip shows the source title and opens the document.
- Optionally designs the visuals using Google Stitch. The wizard validates Stitch's output at every pass to ensure that no citation is paraphrased, summarized, or dropped.
- Hands you a single HTML file you can drop on any free static host — Cloudflare Pages, GitHub Pages, or another service — or open locally to see your finished product.
The wizard and its project files run locally on your computer. Research and optional visual design use the online services you configure.
The public Fractal Framework showroom replays a prepared example of the project's actual workflow. It shows how a location becomes a research question, how sources stay attached to claims, and where a person reviews the result before publication — with no local setup or live model call required.
Most AI-generated content fails for a simple reason: you cannot verify its claims. The writing can sound confident even when facts are paraphrased incorrectly, hallucinated, or linked to sources that do not exist.
Fractal Framework addresses that by:
- Preserving the citation trail. Hash citations from the research stay
in the generated page. When source resolution succeeds, a token like
[WRRC_Mohave_County_Page20_AquiferDeficit]links to a document from the campaign's NotebookLM source set. - Using source quality tags. Government records, established journalism, and weaker sources are marked separately so readers can judge the evidence for themselves.
- Running a validation gate on the visuals. If the visual designer drops or changes a citation while designing the page, that pass is rejected and tried again.
- Keeping a research record you can inspect. Research queries, sources, and responses are logged so someone can retrace how the site was built.
A citation tooltip showing the source document's title, URL, and key.
A core idea of the framework, in plain terms:
In every county in the United States, there's at least one issue area where all five of these are true:
- The problem is real (not invented, not exaggerated).
- The current response is inadequate (regulators are absent, asleep, or outmatched).
- Community awareness is low (most residents don't know about it).
- The regulatory framework is weak (laws don't cover it, or aren't enforced).
- The resources to fix it are within reach (it's not "we need a billion dollars" — it's "we need people to know").
These are activism vacuums. They're the places where one well-built campaign can actually move the needle, because nobody else is paying attention. The framework helps you find one in your own community and build the kind of legible, defensible public-facing site that lets others see what you see.
Current requirement: this is a local source application. You need to download or clone it, install its dependencies, and connect the external research services you want to use.
For Windows, the recommended starting point is the
v1.1.0 release.
Download the runnable Windows source ZIP from Assets, extract it, and
open the extracted folder. Cloning the repository is the alternative if
you use Git or want the current main branch instead of the versioned
release.
- A computer running Windows, macOS, or Linux.
- Node.js 22+ and pnpm.
- Python 3.11+ for the research engine.
- A Google account for the external NotebookLM research service.
- Your own API key for one external LLM provider — Google Gemini, OpenAI, or DeepSeek.
- Optional: a Google Stitch API key, if you want AI-generated per-campaign visual design through that external service.
Open PowerShell in the extracted release folder. If you prefer Git, clone the repository first and run the same commands from the cloned folder.
# Optional alternative to the release ZIP
git clone https://github.com/anitacigawet/fractal-framework.git
Set-Location fractal-framework
# Install the wizard dependencies from the locked versions
Set-Location wizard
pnpm install --frozen-lockfile
# Install the research engine's Python dependencies
Set-Location ../engine/notebooklm_bridge
py -m pip install -r requirements.txt
py -m playwright install chromium
# Authenticate to the external NotebookLM service once
notebooklm login
# Create your local settings file
Set-Location ../../wizard
Copy-Item .env.example .env
# Build and run the wizard
pnpm build
pnpm startOpen http://127.0.0.1:7101 in your browser. Add your API keys via
the Settings page, click "Set up Framework Notebook," and you're ready
to try the "Pick for me" flow with a location.
On Windows, you can instead double-click Launch_Wizard.bat from the release
root after installing the Python requirements and authenticating NotebookLM.
It installs the locked Node dependencies, verifies the bridge, builds the app,
and opens one local wizard process. Run Launch_Wizard.bat /verify from
Command Prompt if you want the checks without opening the wizard.
The generated site preserves hash citations from the research and tries to resolve each one to a source URL. A citation can remain unresolved, and a linked source can still be wrong or misrepresented. Check each claim against its source before publishing.
wizard/— the local React application and its server.engine/— the Python bridge and prompt files used to work with NotebookLM.docs/— the methodology files loaded by the wizard, plus the images shown in this README.Launch_Wizard.bat— the Windows launcher.
For the methodology, start with
docs/fractal_framework_methodology_guide.md.
Fractal Framework was developed using Arizona groundwater campaigns as working examples. The wizard interface includes a human-directed design created with Claude.ai. Optional campaign-specific visual generation uses Google Stitch.
Fractal Framework is source-available under the PolyForm Noncommercial License 1.0.0. You can study, modify, and redistribute it for noncommercial purposes under the license terms. Commercial use is not granted.