FAQ

The product

What does aswritten do?

aswritten lets you install human expertise into AI and verify where every claim comes from. An expert — a practitioner, consultant, author, or your own team — curates their perspective into a living body of knowledge. That perspective gets installed into AI at the moment of need, and every claim the AI makes traces back to the human source.

How is this different from a Claude.md file or project instructions?

A Claude.md file is static instructions you write and maintain by hand. It tells the AI what to do — but it doesn’t install perspective. It provides information, not judgment. aswritten gives your AI the way an expert thinks about a domain — the opinions, the processes, the judgment calls earned through practice. The knowledge graph has conviction levels (how settled is this?), provenance (who said it and when?), and domain structure. When your thinking evolves, the perspective updates and every AI session reflects it. A Claude.md can’t do that.

Why not just use RAG?

RAG retrieves documents. aswritten installs perspective. The difference: RAG gives your AI more information (here are some relevant docs). aswritten gives your AI direction (here’s how this expert actually thinks about this domain, what’s settled vs. in debate, and what judgment calls to make). The knowledge graph isn’t a document store — it’s structured expertise with conviction levels and relationships. Your AI doesn’t just have access to more data; it reasons the way the expert reasons.

How does this prevent AI hallucinations?

Every claim in the perspective traces back to a specific person, in a specific context, at a specific time. When your AI makes a recommendation grounded in installed expertise, it cites the source. When it doesn’t have expertise to draw on, it flags the gap. You can see what’s grounded vs. what’s the AI’s own reasoning. The system doesn’t eliminate hallucination — it makes it visible.

What makes aswritten different from Notion AI / Glean / other knowledge tools?

Those tools search your existing documents. aswritten installs expertise that may never have been documented — the reasoning behind decisions, the methodology that makes your best people effective, the context that new hires spend months absorbing. It structures that knowledge into a perspective that steers AI behavior, not just informs it. And it lives in Git, not a vendor cloud — branchable, reviewable, diffable, yours.

How does the AI “think like an expert”?

When you install a perspective, the AI loads a structured snapshot of that expert’s knowledge — organized by domain with conviction levels that tell the AI what’s settled vs. what’s still in debate. The AI cites this perspective when making recommendations, flags when it’s working outside the expert’s knowledge, and can combine multiple perspectives in a single session. Install a grant writing perspective and a disability-inclusive design perspective together, and your AI reasons across both.

Using it

I was invited to a perspective — what do I do?

Log in with the email from your invite (no sign-up, no card — you’re already on your team’s tier) and follow Quick Start: I was invited. Then ask your AI whatever you’d ask a senior person on the team.

Why isn’t aswritten responding to my prompt?

Most often this is the Claude Desktop Project gotcha: the Project’s custom instructions must contain your perspective’s ASWRITTEN.md/CLAUDE.md content, AND chats must be started from inside the Project — not the home screen. See Quick Start: Troubleshooting for the checklist.

Do I need GitHub?

For most things, no — and this surprises people, so here’s the full picture:

• Using a team’s perspective needs no GitHub at all. Being added to the team is what grants load/ask/remember access. Many team members never touch GitHub.
• Building your own on Expert needs no GitHub account — we manage the repository for you. Optionally connect GitHub for multiple repos and version control access. On Team, you connect your own GitHub org — you own your data.
• PRs: every saved memory arrives as a PR someone reviews and merges — but it doesn’t have to be you. Most teams have a lead who handles it. The AI manages the Git workflow (branches, PRs); the review is your chance to see exactly what the system understood from your words.
• Claude Code with the repo cloned locally is the one workflow that requires repo access, because it reads the files directly.

What makes a good memory?

Direct quotes. The reasoning behind decisions, not just the decisions. Who was involved, what alternatives were considered, and why you chose what you chose. Memories are source material — the extraction pipeline needs primary material to work with. A memory that says “we decided to use PostgreSQL” is weak. A memory that captures why, what you considered, and what tradeoffs you accepted is strong.

This is why your call transcripts are usually the best raw material you already have: they’re full of exactly that — people reasoning out loud, disagreeing, deciding. Paste one in and let your AI draft the memory from it.

How do I know it’s working?

After your first memory merges, load the perspective. Ask a question about your project — if the answer cites your decisions with provenance instead of giving generic advice, it’s working. Use introspect to see coverage by domain and identify where knowledge is thin.

Pricing

What’s the difference between Team and Organization?

Scope and privacy. On Team, your team collaborates on shared perspectives — live, not static snapshots. You own your repos and can publish to the marketplace. On Organization, you create a private catalogue for your organization — role perspectives, department expertise, onboarding frameworks — all internal, all auditable. We create the perspectives for you: interview your senior people, process existing docs, and deliver working perspectives. Organization also includes agent deployment, on-prem, dedicated support, and unlimited seats. See Pricing for the full comparison.

Can I use this on-prem? (HIPAA, GDPR, compliance)

Yes. On-prem deployment is available on the Organization tier when compliance requires it. The architecture separates deterministic operations (commit, assemble — these run locally or in your infrastructure) from LLM operations (remember, introspect, generate — these route through your approved providers via your own API keys). Zero data residency: nothing leaves your network. This covers HIPAA, GDPR, data sovereignty, and ISO 27001 requirements.

Technical / architecture

How is git-native different from model-held context?

Model-held context (like Copilot Memories or Claude’s memory) is scoped to one tool, one user, one conversation thread. It’s not version-controlled, not reviewable, and it disappears if you switch tools. Git-native means your expertise is code — branchable, reviewable, diffable, portable across every AI tool. Multiple people collaborate on how your AI thinks through the same workflows they already know.

What is “perspective vs. information”?

Information is facts — what search and RAG provide. Perspective is how to interpret, frame, and apply information — the opinions earned through practice, the judgment calls, the sense for when the conventional wisdom is wrong. Wisdom is knowing when and how to apply perspective in context. Every existing AI tool operates at the information layer. aswritten operates at the perspective layer. No amount of better retrieval turns information into perspective. They are different things.

What is the extraction pipeline? How do memories become knowledge?

1. You save a memory via the remember tool
2. Extraction runs synchronously inside the tool call (~2-3 minutes)
3. An LLM reads the memory and produces SPARQL transactions — structured knowledge statements (entities, relationships, conviction levels)
4. Transactions are validated against the ontology
5. Memory and TX files are committed together atomically — if extraction fails, nothing is committed
6. A PR opens showing exactly what was extracted — you can review every fact
7. You merge; the perspective updates on next load

Every fact in the graph traces back to the memory it came from.

How does the ontology work? Can I customize it?

The ontology defines entity types (Actors, Claims, Decisions, Narratives), relationships, and conviction levels. It’s an RDF schema — the knowledge graph is structured data, not free text. The base ontology covers common knowledge patterns and has proven domain-agnostic across organizational strategy, clinical practice, engineering, and consulting contexts without modification. Custom ontology extensions are available on the Organization tier for domain-specific needs.

How does review mode work? What are lenses?

When a memory is extracted and a PR opens, the review process uses an optometrist model — iterative convergence through shifts. The system identifies domains affected by the new knowledge and generates perspective shifts: “Better through lens A or lens B?” You converge on the right interpretation through comparison, not by evaluating claims in isolation. This catches misinterpretations before they enter the perspective.

What does “underneath the text” mean?

Every document — a blog post, a pitch deck, a status report — has expertise underneath it. aswritten makes that layer explicit. When you generate content from installed expertise, every factual claim can be annotated: grounded (traces to a specific memory from a specific person) or ungrounded (gap in knowledge). The cite tool maps any text against the knowledge graph and reports what’s supported vs. what’s asserted without provenance. The text is the surface; the graph is what’s underneath.