Authorship AI: A multi-agent writing coach that strengthens users’ cognitive autonomy

My Cosmos Institute AI x Truth-Seeking grant competition final write-up. Sad it's over, but it's just the beginning for AI products that expand human agency.

I first felt the AI frontier getting weird when I saw my career get automated overnight. I had been working in writing-intensive roles: business journalism, marketing and communications, and market research, and my mode of output, the craft I’d been honing for years, evolved overnight from the long-form article to the prompt.

In my view, writing is thinking. Yet I’ve too often heard the argument that writing is inessential work — a non-technical side quest at best. Writing is apparently just for essays in K-12 and university humanities education, and for media and media relations professionals, and therefore it’s too small a problem to address. I beg to differ.

If you comment on this post, you are writing. If you draft a memo to your manager, or a product requirements document, you are writing. If you draft an email, you are writing. If you post on LinkedIn, you are writing. If you apply for a job, you are writing. If you brainstorm with bullet-point ideas and a note-taking app or whiteboard, you are writing. And to execute all these workflows, you must think and determine how to express your thinking, which is writing.

Some might argue that AI writing tools can help us generate ideas more efficiently. However, my concern with a world of synthetic content generation — where we zero- or single-shot prompt LLMs to full-length drafts — is that we might lose our ability to think critically and develop original ideas if we rely too heavily on these tools.

Today’s LLMs are showing early signs of powerful reasoning abilities; they fill in many of the blanks and connect dots that we don’t wish to, or struggle to, or claim we are to busy to connect. When we use ChatGPT, Claude, Gemini, or any wrapper tool to write for us, we generally permit them to think for us, too.

MIT Media Lab researchers found weakened neural connectivity in a brain scan study on the use of LLMs in essay-writing. It’s a finding that aligns with many of my user research (UXR) interviewees; when I spoke with K-12 educators, higher education professors, and private sector learning and development (L&D) leaders, I heard the same observation across the board: students and professionals alike are misrepresenting AI-generated work as their own, losing out on critical thinking and authentic expression in the process, and sometimes even can’t define words and ideas in their final deliverables. It’s becoming harder than ever not just to detect or challenge AI-generated content, but to incentivize intellectual capacity-building in the era of AI assistance.

Over the long term, this has profound implications for gradual disempowerment, a scenario where an uptick in AI capabilities risks a loss of human influence in the world, so pernicious that humans don’t notice their displacement with AI systems.

That’s why I decided to build Authorship.

I believe we can build AI tools that not only aid in productivity but also strengthen our cognitive abilities. Authorship is a writing coach that helps users think for themselves, express their ideas more effectively, and ultimately become better writers and thinkers. In this essay, I’ll share the journey of building Authorship and the principles that guide its design.

The AI writing coach that didn’t exist

I wanted to build from first principles the writing coach that didn’t exist: one that strengthened the user’s original voice and ideas, and in doing so, yielded higher-quality outputs than ChatGPT or Claude alone. Imagine an AI writing tool that educators and workforces deemed “kosher” for use, where users would leave with better-expressed ideas that sound like them and appropriately influence their intended audience. Authenticity, not sterility.

Before I knew what I was getting into, I declared my intention to build an MVP with these features:

• Voice preservation: Models trained on your corpus to spot drift and nudge you back to your style.

• Socratic scaffolding: Guided prompts that ask you to outline claims, evidence, and counter-arguments before the model completes anything.

• Browser extension: Delightful user interface that gives feedback on your prompting and your drafts.

• MCP server connection: Highly contextualized outlining and writing guidance that requires regular writing sample submissions to augment your unique capabilities.

The final MVP focused more heavily on the first two features, saving the last two for production-grade future directions. I kept our North Star focus on the question, how can I build an AI tool that helps users think for themselves and sound like themselves in a user-friendly way?

There were two key tinkering phases for the Authorship project, which I’ll summarize and expand on below

1. Foundational UX and system prompt

2. Multi-agent architecture

I am grateful for tremendous engineering help from Andy Neuschatz (my former colleague from The Information’s engineering team) in the first phase, and John Lund (a philosopher-builder I met via the BlueDot Impact community) in the second phase.

Andy built the foundational architecture: the open-source editor, the API connections, the infrastructure that made everything else possible. When the project pivoted toward fine-tuning and multi-agent orchestration, I brought on John Lund, whose background in AI engineering fit the new technical direction.

Phase One: The single-agent trap

Early on, I built what I thought was clever: a sophisticated system prompt that instructed Claude to monitor for cognitive dependency, adapt its response style, withhold capabilities when users over-relied, and preserve their voice.

Andy and I met biweekly and collaborated on the early vision: An open-source word processor with an AI sidebar: about 1/4 the browser window would show the Authorship assistant, and the remaining 3/4 would be whitespace for the user to draft. Authorship would periodically send updates from the whitespace to the API and produce valuable feedback, and the user could have a native chat conversation with Authorship on-demand.

Andy built the word processor and mission-critical project management scaffolding, and I built and tested a prompt library in the Claude Developer Console until I found the optimal system prompt for Authorship.

It worked. The outputs were genuinely different from vanilla Claude — more Socratic, more restrained, more focused on the user’s own thinking.

But after months of tinkering, I had a sinking feeling: had I just built a wrapper? A really good prompt is still just a prompt. I could demo it, but what had I actually contributed?

I wrote to a Cosmos advisor with my concern, and his feedback guided me toward using my remaining time and resources to aim higher with technical ambition, and to product-manage the vision for an AI-native engineer.

He was right. A prompt can shape one interaction, but it can’t specialize. It can’t learn. It can’t break down a complex problem into constituent parts and address each with domain-expertise. That’s when the project entered its second phase.

The breakthrough came from a conversation with Roger Thompson, who directs writing programs across Arizona State University, and to whom John introduced me. He explained that most writing instruction fails because it conflates three distinct skills that are actually orthogonal:

• Rhetorical awareness: matching your message to your audience. Who are you writing for? What do they already believe? What would move them?

• Epistemic development: constructing and justifying claims. Are you asserting something as fact, attributing it to a source, or evaluating evidence? How do you know what you claim to know?

• Authentic voice: the choices that make writing sound like a particular person rather than generic prose. Where do you commit versus hedge? Where do you show up versus disappear into safe, impersonal language?

You can have strong audience awareness but weak reasoning. You can reason well but write in a voice indistinguishable from any other competent professional. Most AI writing tools collapse all three into a single “improve my writing” function — which means they can’t help you develop any of them deliberately.

Phase Two: Three agents, hard constraints

In the second phase of the project, I decided to build three specialized agents, each focused on a different aspect of writing instruction. I fine-tuned each one on the Mistral API, with tailored prompts and training data examples, to help users develop specific skills. We felt that the reasoning capabilities and inference latency in Mistral Small relative to token price were sufficient for proving the concept of our three agents:

1. Audience Coach handles rhetorical awareness. It’s grounded in Aristotle’s Rhetoric treatise — the original framework for matching message to audience. It asks questions like: Who are you writing for? Does your tone fit that community? Are you meeting your audience where they are? Do you appropriately and effectively stir emotion in that community? I found that this approach helped users think more carefully about their audience and tailor their writing accordingly.

2. Reasoning Coach handles epistemic development. It’s grounded in research on epistemic stances — how learners develop from seeking “the right answer” (absolutist) through “all opinions are valid” (multiplist) to evaluating evidence and making justified claims (evaluativist). The Coach only asks questions. If you say “What should I write?” it responds “What claim are you trying to make?” If you cite a single source as proof, it asks “What would someone who disagrees point to?”

3. Voice Coach handles authenticity. It reflects rather than advises or improves. It notices tendencies like redundancy and variety, like where you sound like a specific person versus where you disappear into generic language, where you assert directly versus attribute to others versus hedge. Then it asks a single question: “Is that intentional?”

I inverted Anthropic’s approach in what I term “constitutional withholding”—a method to ensure AI doesn’t just generate answers, but prompts users to interrogate their own reasoning first.

In Phase I, I found that constitutional AI models like Claude Sonnet 4.5 can be “constitutionally modified” to withhold rather than maximize capabilities—creating productive yet lightweight friction that preserves user agency. In Phase II, I trained the three agents to double down on this for the user’s benefit.

Authorship’s constitutional withholding approach brings AI for truth-seeking to life by promoting a habit of interrogation and self-reflection in a user-friendly interface.

Instead of training models to refuse harmful requests, we’re training them to withhold helpful capabilities when providing them would undermine cognitive development. The constraints are built into each agent’s core directives, rather than added as afterthoughts. Reasoning Coach cannot give you the answer; it’s not refusing to be helpful; answering isn’t within its capability set. Voice Coach cannot tell you what your voice should be. Each agent has one job and hard boundaries.

The stakes go beyond education. If humans can’t independently construct arguments or spot flawed reasoning, we can’t meaningfully oversee AI systems. Cognitive dependency is a civilizational risk to our ability to govern powerful AI; not just a humanities pedagogical concern. (It’s also why high-quality writing instruction remains vital in the AI era.)

This mirrors the role of a good human writing instructor. The best teachers ask questions that help you see what you’re doing, rather than fix your work for you. They create what educational researchers call “productive struggle”: difficulty that leads to growth. AI that removes all friction removes the conditions for learning.

What we built

When I tested the prototype, I noticed something surprising: running all three agents at once forces users to confront their blind spots in real time.

Authorship runs all three agents in parallel, then synthesizes their perspectives into a single, prioritized summary response. They see what matters most right now, with the option to drill into any agent’s full analysis.

Early testing suggests the approach is valuable. Engaging with the Reasoning Coach produces more nuanced claims. Voice Coach observations help users start noticing their own patterns. Audience Coach helps users get their message across. The feedback loop strengthens their writing and their thinking, which was the whole point.

Authorship has soft interest to run a university-wide pilot in 2026, comparing learning outcomes between students using Authorship versus standard LLMs. If we can show that constitutional withholding produces better epistemic development — not just better text — that’s a meaningful contribution to how we design AI for education and knowledge work.

Future directions for Authorship

The current prototype surfaces feedback from all three agents, but the next iteration will be smarter about prioritization. We’re building an orchestration layer that analyzes the user’s text and delivers only the 1-3 most critical insights — not three parallel feedback streams. The goal is a focused, numbered list of observations that users can respond to quickly: Imagine, “1. Your audience isn’t clear. 2. You’re hedging your main claim. Which do you want to address first?” This gives users an intuitive, short-form way to engage with Authorship and make targeted improvements without cognitive overload.

We’re also investing in onboarding and persistence. Right now, each session starts fresh. Future versions will store voice fingerprints and writing history, so Authorship learns your patterns over time and can detect drift across documents, not just within a single draft. A polished, private onboarding flow will help users articulate their goals upfront: What are you writing? Who is it for? What’s your relationship to this material? These inputs will route feedback more intelligently from the start.

Toward an intentional steering of the wheel

I started this project because I watched my own craft get automated and felt something important slipping away. Not the job — jobs change — but the thinking that the job required. The struggle to articulate an idea is the process of understanding it. When we outsource that struggle, we lose the cognitive development that produces genuine understanding.

Authorship is a bet that we can build AI differently. Not AI that writes for you, but AI that helps you write better. Not AI that thinks for you, but AI that strengthens your capacity to think. The technology exists to do either. The question is what we choose to build.

We’re already on the path. It’s fast-moving. And I’d rather steer it than watch it go by.

Author’s Note: I used Authorship to revise this draft.