The Letter Assistant is a free web tool, built and hosted by Matt Ordeshook, that walks a
California renter through a tenant-rights problem in plain conversation and hands them back
a completed, ready-to-mail letter, built from Tenants Together's own catalog
of 33 sample letters. It runs at a public URL, needs no login, and produces a printable PDF
the tenant can download or email to themselves in one sitting.
01
How a tenant experiences it
No account, no forms up front. Just a conversation that ends in a document.
1
Say what's happening
The tenant opens the page and describes the problem in their own words, or taps a starter like My landlord kept part of my security deposit. The assistant leads with a disclaimer: it isn't a lawyer, this isn't legal advice, and it asks a couple of short, specific questions to understand the situation.
2
It names the right and proposes a letter
Once it understands the problem, it explains the relevant California tenant right in plain language and names the one letter from Tenants Together's catalog that fits, and lets the tenant confirm before committing to it.
3
It collects the specifics, a few at a time
Move-in date, deposit amount, the landlord's name and address: whatever that specific letter needs. The assistant asks conversationally rather than dropping a form on the tenant, and a checklist panel on the side fills in as fields are captured.
4
Download or email the finished letter
Once every required field is filled, the tenant can preview the letter, Download PDF, or have it emailed to themselves. Every letter closes with the same reminder: send it by certified mail, keep a copy, and this is a template, not legal advice.
Where the AI sits in the pipeline, and just as important, what it's not allowed to do.
BrowserChat pageThe tenant's messages
→
Claude (Anthropic)Reads & guidesUnderstands the problem, asks questions, picks the letter
→
Server codeFills & rendersSlots answers into the real letter, builds the PDF
→
TenantPDF letterDownload or emailed copy
letter textEvery word of every letter comes from Tenants Together's own 33 PDFs, transcribed once into the system. Claude never writes letter language from scratch; it only picks which letter fits and fills in blanks.
conversationStored in a private database (Supabase) tied to that browser session only, so the assistant remembers context if the tenant navigates away and comes back.
the PDF itselfAssembled by ordinary code (a PDF layout library), not generated or "written" by the AI, so formatting is identical every time.
The model in use is Claude Sonnet,
made by Anthropic, the same family of AI behind Claude.ai. It's given the full Tenants Together
"Know Your Rights" reference and the letter catalog as background reading before every conversation,
and it can only take three actions: choose a letter, save a field the tenant just gave it,
and clear a field to redo it. It cannot invent a letter that isn't in the catalog, and it
never generates the PDF itself; that's fixed, testable code, so the letter a tenant receives is
always exactly what Tenants Together published.
Never invents statutes or deadlines
Instructed to say "I don't know" and refer to the hotline rather than guess at law it isn't grounded in.
Leads with court deadlines
If a tenant mentions an actual eviction lawsuit (a Summons & Complaint), it stops and flags the ~10-business-day court deadline before anything else.
One letter, one topic
Refuses to blend two situations into one letter; if two could apply, it explains the difference and lets the tenant choose.
Never asks for sensitive data
No Social Security number, immigration status, or diagnosis details, even for disability-accommodation letters.
03
What it costs to run
Usage-based, not subscription-based: the bill scales with how many tenants actually use it.
Component
What it does
Typical cost
Claude (Anthropic API)
Reads the conversation, asks questions, picks the letter
≈ 10–25¢ per completed letter
Hosting (Vercel)
Serves the page and runs the server code
$0 incremental*
Database (Supabase)
Stores in-progress conversations
$0 at current volume*
Email delivery (Resend)
Sends the PDF when a tenant asks to have it emailed
$0 at current volume*
All-in, per completed letter
Rounded up for headroom
≈ $0.10 – $0.30
*Riding on infrastructure already in place for the rest of the site, inside each service's free tier
at this volume. The only cost that scales directly with usage is the Claude API call, so 1,000
completed letters in a month runs roughly $100–$300, not a fixed monthly fee regardless of use.
Cost overview: Nothing to worry about during testing. The $15 already put
in comfortably covers this build-and-test phase, running on Matt's own hosting and Anthropic
account. Once it's ready for the public, we'll just have a quick chat about its permanent home
(likely Tenants Together's own domain and Anthropic account) and what traffic to expect.
04
Privacy: built to hold as little as possible
No accounts, no tracking, and nothing kept longer than it takes to produce the letter.
Auto-deleted within 3 hours
Any conversation that goes quiet longer than that — a closed tab, a dead phone — is wiped automatically and permanently. No exceptions, no manual cleanup needed.
One-click permanent delete
A "Clear my data" button in the assistant deletes a conversation immediately, any time, without waiting on the 3-hour window.
Anthropic doesn't train on it
Per Claude's commercial API terms, conversations sent through the API aren't used to train Anthropic's models.
Letters aren't stored
Each PDF is assembled fresh from ordinary code and streamed straight to the tenant; the finished file never touches our servers.
No accounts, no tracking
No sign-up, no cookies, no analytics pixels anywhere on this tool.
Session-only by default
Closing the tab clears a tenant's local copy right away; the server-side copy follows automatically within the 3-hour window above.
Ideas on the table for a future version. None of this is built yet, and nothing here gets decided without Tenants Together.
Language
A Spanish version of the assistant
About half the letter catalog already has an official Spanish translation sitting in Tenants Together's own source PDFs, so that half could ship first. The rest would need Tenants Together to supply or review a translation before it's safe to publish. Claude itself needs no extra work to hold up its end; it's natively fluent in Spanish.
Cost
Fewer AI calls for the routine, fill-in-the-blank parts
For a tenant who taps a starter button and answers straightforward questions (a name, a date, an amount), the assistant could collect those directly with ordinary code instead of asking Claude to ask each one. The AI would still handle the parts that need judgment: explaining the right, understanding an unusual situation, picking the correct letter. Cutting AI calls out of the routine path lowers the cost per completed letter more than any other single change.
Content
More than just the letter in the PDF
The generated PDF could carry a page or two beyond the letter itself: a sponsorship or funder acknowledgment page, a resources page (the hotline, the attorney directory, other tenant assistance programs), or fuller instructions on how to send the letter and what to do if the landlord doesn't respond.