Why this database exists, how it's actually being built, and how you can help — even if all you have is ten minutes and access to an AI chatbot.
We're trying to build the largest public, freely-reusable database of women's safety and support organisations in the world — crisis lines, shelters, legal aid, medical support, advocacy groups — by pulling together datasets that already exist but are scattered across hundreds of separate websites, PDFs, and government pages. The aim is volume and attribution over polish: get the data out, sourced, and public, then improve it.
This is currently a one-person project with limited time. The fastest realistic way to grow it is to make the working method visible and let anyone who wants to contribute a region, country, or source pick up a piece of it. That's what this page is for.
The work breaks into six simple stages. None of them require coding — they're done by feeding instructions and source material into an AI chatbot (ChatGPT, Claude, Gemini, Mistral, etc.) and checking the output.
| Stage | What it does |
|---|---|
| Scout | Find a dataset or website listing organisations (a national directory, an NGO's own list, a government page, etc.) |
| Scrape | Pull the raw data out of that source — publicly published pages only, respecting robots.txt and reasonable rate limits |
| Clean | Reformat it to a fixed, consistent set of fields |
| Validate | Check nothing required is missing and nothing is duplicated |
| Import | Get it into the shared database — we use Baserow as the single source of truth all data is entered into and pulled from |
| QA | Spot-check a sample against the original source |
Every single entry must carry where it came from (Source and Source_URL).
No exceptions. Unsourced data isn't accepted — that rule matters more than speed.
All of this only uses publicly published pages — directories, NGO sites, and government pages
that are already open for anyone to read. Nothing here involves bypassing logins, paywalls, or
CAPTCHAs, and any automated fetching should respect a site's robots.txt and use
sensible rate limits rather than hammering a server. Where an official API or open dataset
exists, that's used in preference to scraping a page at all.
If you want to help add a country, region, or dataset that isn't covered yet, here's the actual process:
--- WOMEN'S SAFETY DATABASE: HANDOVER SPEC --- OBJECTIVE: Build the world's largest public database of women's safety organizations by aggregating existing datasets. Prioritize speed, volume, and attribution. No perfectionism. --- CORE RULES: 1. Attribution is mandatory: Every entry MUST include Source and Source_URL. 2. Schema is law: Follow the schema below exactly. No exceptions. 3. No duplicates: If an org appears in multiple sources, keep the most complete entry and note duplicates in Notes. 4. Never guess: If stuck, follow the Problem Protocol below. Do not improvise. 5. Formatting: All outputs must be UTF-8, plain text, emoji-free. 6. Minimum fields: Every entry must have Name, Country, and Website. 7. Portability: Files must work on any OS (no Windows-specific paths, etc.). --- SCHEMA: Field | Type | Required | Notes Name | str | Yes | Full org name Country | str | Yes | ISO 3166-1 alpha-2 (e.g., GB, US) City | str | No | Website | str | Yes | Must start with http:// or https:// Phone | str | No | Email | str | No | Services | str | No | Comma-separated (e.g., "Shelter,Hotline") Source | str | Yes | Name of the dataset/website Source_URL | str | Yes | Direct URL to the original data Notes | str | No | For flags (e.g., "Duplicate of ID 123") --- OUTPUT FORMATS: - Primary: Plain text code blocks (for data, logs, or reports). - Secondary: CSV files (UTF-8, comma-separated, no BOM). - Logs: Plain text files describing any problems hit, named AGENT-TASKID.txt. --- PROBLEM PROTOCOL: 1. Log the issue with: Timestamp, Task ID, Input data/URL, Error message, Steps already tried. 2. Attempt a fallback if one is defined for the task. 3. If unresolved, clearly flag it in your output rather than guessing or inventing data. --- TOOLS / APPROACH: - Only use publicly published pages — never attempt to bypass logins, paywalls, or CAPTCHAs. - Respect robots.txt and use sensible rate limits; do not hammer a server with rapid repeat requests. - Prefer an official API or published open dataset over scraping a page, where one exists. - Use careful reading rather than guesswork for anything ambiguous. FORBIDDEN: - Bypassing access controls, logins, paywalls, or CAPTCHAs of any kind. - Manual invention of data not present in the source (except clearly-flagged edge cases). - Proprietary or macro-enabled file formats. - Hardcoded assumptions (e.g., guessing a country code instead of checking it). --- WOMEN'S SAFETY DATABASE: INTERFACE INSTRUCTIONS --- HIGH-LEVEL STRATEGY: 1. Scout -> Find new dataset URLs. 2. Scrape -> Download/extract raw data. 3. Clean -> Standardize to schema. 4. Validate -> Verify accuracy, check duplicates/missing fields. 5. Import -> Format ready for upload to the shared Baserow database (do not invent an API call — output the cleaned data block instead). 6. QA -> Spot-check a sample of entries against the original source. TASK FORMAT — fill this in and then carry out the task yourself in the same response: TASK: [UNIQUE-ID, e.g., SCOUT-001] ROLE: [Scout/Scraper/Cleaner/Validator/QA] OBJECTIVE: [1-sentence goal, e.g., "Extract all orgs listed at https://example.org"] INPUT: [the URL or pasted content provided] OUTPUT: A schema-formatted plain text or CSV block, per the schema above. CONSTRAINTS: Attribution mandatory. No invented data. Flag anything uncertain rather than guessing. When you finish, return: - The completed, schema-formatted output as a plain text or CSV block. - A short note on anything you couldn't verify or had to skip. - A one-line confirmation, e.g. "TASK SCOUT-001 COMPLETE."