We Open-Sourced the Guardrails Inside GoldenDoodle AI

Pull up the last few emails you got from organizations you support. You can probably pick out the ones AI wrote by the second paragraph. An em dash where a comma would do. A program that “empowers” people on a “journey.” A closing line that repeats the opening.

A year ago that was a hunch. Now you can do it without trying. In a February 2026 survey of 346 nonprofits by Virtuous and Fundraising.AI, 92% said they use AI in some capacity, and 81% said they use it ad hoc, one person at a time, with no shared way of working. Nonprofits were slower to adopt AI than most sectors. The ones that did adopt it are now publishing with it every week, and readers have learned the sound.

Why the sound matters

If your organization wants to be the place people turn to, sounding like every other organization costs you something. The AI habits become the thing people notice. A reader who spots them starts thinking about the tool instead of what you’re offering. When people hear AI, they stop hearing you.

For someone who has lived through something hard, the cost is higher. They reached out to a person. What came back reads like it was assembled by a system, polished and a little too sure of itself. It can feel like being managed by a machine at the exact moment they needed to be met by a human being. Trauma-informed practice asks organizations to build safety and trust. Writing that sounds automated works against both, even when every fact in it is correct.

What we’re giving away

GoldenDoodle AI started as a writing studio for organizations whose readers may have lived through trauma. As far as we know, it was the first AI writing tool built for organizations that was trauma-informed by default. Nobody had to switch the standard on. Every draft ran through it. Its image generation is still the only AI image generation we’ve found that’s held to a trauma-informed standard.

Those guardrails lived inside one app, and they helped the teams who used it. Starting today, the writing guardrails are free for the AI you already use, whether that’s Claude, ChatGPT, Gemini or an agent:

Human-centered language for anyone who writes about people. Person-first language, people described by what they do instead of what they lack, and the person at the center of their own story with the organization in a supporting role.
Trauma-informed language for organizations whose readers may have lived through trauma. It builds on the human-centered standard and adds SAMHSA’s trauma-informed principles, plus practices for fundraising, storytelling and consent.
Human voice for anyone who doesn’t want their writing to read like AI. It takes out the habits readers have learned to spot.

All three are free, with no account or sign-up, and they’re yours to change. Open source means the files are public and anyone can use them. The license we chose, CC BY 4.0, lets your organization edit them to fit your own voice, programs and community, and share your version with partners, as long as you credit GoldenDoodle AI. If you find something missing, tell us. The skills page walks you through setup in about two minutes, and you don’t need to be technical.

Why we test them every month

Plenty of people publish skill files now. What a file can’t tell you is whether it still works after the model underneath it changes, and the labs update their models every few weeks. A skill that worked in September can read differently to the model that shipped in November.

So every month, the Human-Centered Language Benchmark runs these skills on the latest models from the major AI labs and the leading open-weight models, and publishes what happened. Since October 2026 we’ve tested them on 8,370 drafts from 17 AI models.

In the November run, 16 AI models on their own got people right 72 times in 100. With the human-centered skill loaded, 99. On writing for readers who’ve been through something hard, the models got people right 63 times in 100 on their own and 97 with the trauma-informed skill. The human voice skill took em dashes from 6.1 per 1,000 words to zero.

That testing is why these files will change. When a new model starts getting something wrong, the standard gets a guardrail for it, and every change is logged with its reason and its source. If you load a skill today, check back. The version that works best with next spring’s models may not match this one.

The skills don’t make AI perfect. With the trauma-informed skill loaded, 3 drafts in 100 still got people wrong in November. Someone on your team should still read before anything goes out.

The studio is still here

GoldenDoodle AI Studio runs on the same standards it always has. Every draft treats the people in it as people. The person is the subject of their story and the organization is the supporting cast. Strengths come before deficits, plain language comes before jargon, and nobody’s story gets told without their say. If your team wants a writing and image tool built around those commitments, the studio is open, with plans starting at $20 per seat.

Start with your own website

Most organizations have never checked how their website writes about the people they serve. Dignity Audit does that for free. Enter your URL at dignityaudit.com and it reviews five core pages against the trauma-informed standard, looking for stigmatizing labels, deficit framing, stories where the organization is the hero, and places where a strengths-based line would serve readers better. Results arrive by email with specific guidance on what to revise and why. If you want every page checked, a full site audit is $29.

If you’re not sure where to start, run the audit. It takes a few minutes and shows you how your website writes about people today.

Scott Eggert, founder, GoldenDoodle AI