Howard

Asking Howard from Claude works again

If you connected Howard to Claude, asking a full question was failing with a connection error even though Howard had finished the answer. Howard now keeps the line open while it works, so long answers and hard questions come back to Claude instead of timing out.

Smarter answers for hard questions, and the answer up front

Today was about getting the right answer the first time, and putting it where you can see it.

• Howard now uses a more capable model for every answer, so the right details from FASRG, EDGAR, and TEA guidance should be there the first time. Difficult questions may take a little longer while it works through the sources.
• When Howard double checks an answer and finds a problem, the corrected version is written by the same stronger model thinking its hardest.
• Answers now start with the direct answer to your question (the yes or no, the code, the screen name). Conditions, steps, and caveats follow underneath, so you can stop reading once you have what you need.

Faster answers, a second set of eyes on account codes, and no more waiting forever

Three changes today. Howard now uses a much faster model for most answers. Because speed should not cost accuracy, Howard also gained a checker that re-reads the source documents after each answer and double checks every account code it mentions. And Howard now keeps an eye on the clock, so a slow question gives you what it has instead of leaving you staring at a spinner until the page gives up.

• While the checker works, you will see a small purple bubble and a light underline on each code it is verifying.
• When it finishes, a note shows which codes checked out. If a code claim could not be confirmed from the sources, Howard automatically asks a more careful model to write a corrected answer, which appears right below the first one.
• Your history now shows whether the codes in an answer were verified.
• If an answer is taking too long, Howard stops waiting. If it had already started writing, you get what it wrote with a note saying it was cut short. If it had not written anything yet, a quicker model answers from the sources already found.
• Asking again, or asking more narrowly, usually gets a full answer after a cutoff.

Fixed API keys and corrected ChatGPT setup steps

Two fixes today: a broken settings page and clearer ChatGPT setup instructions.

• Creating an API key on the Connect Howard to your AI assistant page was silently broken. Clicking Create did nothing. That is fixed now, so creating and revoking keys works again.
• Our instructions for connecting Howard to ChatGPT pointed at a menu that no longer exists. They now match what ChatGPT actually looks like today, including which plans support this. We also suggest naming the connection Howard, because then you can just say ask Howard about something and your assistant will know what you mean. Howard's tools are marked read only too, so ChatGPT stops asking you to approve every single lookup.

Email and rate limit fixes, and connect Howard to Claude or ChatGPT

A few things today. Howard's emails now come from [email protected] and are sent through a dedicated email service, because some school districts were quietly filtering the old messages. Everyone was also sharing a single limit on how many questions could be asked per minute, so a few people using Howard at once could cause others to be told to slow down for no reason. And you can now let your AI assistant look things up in Howard's library for you. Go to Settings and create an API key, then paste the web address it gives you into Claude or ChatGPT. Your assistant can search Howard's documents and get exact citations without you leaving the chat.

• If you signed up and never received your verification email, try again. It should arrive now.
• Check your junk folder if you still do not see it, and ask your technology department to allow mail from schoolbusinessbot.com.
• Each person now has their own question limit instead of sharing one with everybody else.
• See the new setup guide for step by step instructions on connecting Howard to Claude or ChatGPT.

Faster loading on school networks, and safer answers

Two things today. Howard now loads properly on locked down school and district networks, and answers are checked before they are shown to make sure nothing unexpected can run in your browser.

• If you are on a school or district network, Howard could take almost a minute to load, and the Live Feed could say there are no questions yet when there really were plenty. Howard was borrowing a small piece of code from an outside website that some school filters quietly block. That piece now comes from Howard itself.
• Answers shown on the Live Feed, your history, and shared links are now cleaned before they appear, so only normal formatting like headings, tables and links can show up. Links to source documents are checked too.
• Howard also sends pages in a more compact form now, which helps most on a slow connection.
• Admins now have two switches on the Live Feed. One shows questions an admin removed from the feed, the other shows questions people chose to keep off it. Both are off unless an admin turns them on, and neither is available to anyone else. Marking a question private still keeps it off the public feed for everyone, and an admin could already see those in your history.

Better web lookups, and easier reading while Howard types

Improvements to how Howard looks things up on the web when its own documents come up short, plus two fixes for reading a long answer.

• Howard now rewrites your question into proper school finance wording before it searches. Asking what is the deal with PIC 37 used to search those exact words, which brought back social media posts. Now it searches for the program intent code by name and finds the real TEA pages.
• Web results are limited to official sources such as TEA, the federal register, and the U.S. Department of Education. A copy of a TEA document posted on some other district's website no longer counts.
• Howard now reads the full page for the top few results instead of only the short preview text, so it has more to work with when it answers.
• You can now scroll up to re-read part of an answer while Howard is still typing. The page stays where you put it instead of snapping back to the bottom on every word. A Jump to latest button appears so you can catch back up, and scrolling to the bottom yourself starts the automatic follow again.
• When Howard reconsiders and writes a revised answer, the first answer is now greyed out and labeled First answer, so it is clear which one is the current one.

Progress updates while Howard works, and no more silent failures

While Howard is working out a hard question it now keeps you posted the whole time: short notes about what it is checking while it thinks, and a running count of how long it has been at it. This works for guests too. Under the hood, a hiccup on the first model now retries that same model instead of quietly switching to a faster, less careful one, and any question that fails to get an answer is now recorded so we can spot and fix the cause.

A better brain, a window into Howard's work, and fewer failures

A big day. Today was all about three things: smarter AI models behind every answer, letting you watch Howard work instead of staring at a spinner, and fixing the ways long questions used to fail.

• Howard now runs on AI models that reason through a question before writing. Answers stick closer to the source documents, with exact citations. Hard questions take longer, and that is on purpose.
• Answers appear as Howard writes them instead of all at once. If Howard reconsiders and digs deeper, the first answer stays on screen and the revised one appears beneath it.
• The status line now says what Howard is actually doing: the search angles it chose, which documents the best material came from, and why it is taking a second pass. It all stays in the Show steps list under the answer.
• On harder questions you will see short notes about what Howard is checking while it thinks, such as which fund a salary belongs in. These are a preview of its working, not the answer.
• Questions that took more than about two minutes were failing with a connection error. Fixed, including for people who are not signed in.
• Complex answers that stopped mid sentence: the length limit is now twice as high.
• Every page has the same header and menu, so Tools, What's New, the live feed and your account are one click away from anywhere.
• The thumbs up and "this answer is wrong" buttons now show up on past answers too, in your History and in reopened conversations.

New: TEA's federal grant monitoring guide (published version)

Added the official published version of TEA's guide for monitoring federal grant money. This is the rulebook TEA uses when it reviews how districts spend federal funds, including what they look for on time and effort paperwork. The earlier draft we had was for training only and is now retired, so answers here match what TEA actually enforces.

More plain-English guides on tricky topics

The knowledge base grew again, covering questions that come up a lot: which fund food service money goes in, when you can buy equipment with federal funds, which costs are allowed under Title II Part A, and a new break down of how the state's extra spending money for districts (ASAHE) works. Each new topic means fewer dead ends when you ask.

Upgraded to the newest edition of FASRG

Howard now answers from FASRG v20, the newest edition of Texas's school accounting rulebook. Old v19 pages are kept out of the way so you won't get stale rules by accident. We also fixed every broken link in the source list and added the current version of TEA's General and Fiscal Guidelines, plus the newest set of grant program guides.

New tool: School Finance Quiz

There's a new tool in the Tools menu: a practice quiz built from the same rules Howard answers from. It's a painless way to check what you actually know about school finance. Your Howard account works there automatically, no separate sign up.

Sharper search results

Howard now prefers the newest version of a document when several editions could answer your question, and warns you when a source is out of date. A long tuning effort also made sure popular pages don't crowd out the exact paragraph you need. Retired and duplicate documents were cleaned out of the library.

A friendlier mobile layout

On phones, the top menu now folds into a tidy slide-in drawer, and Howard got a proper little icon in your browser tab.

Share any Q&A with a link

You can now turn any conversation into a shareable link, great for sending an answer to a colleague. Visitors can also try Howard without an account and see more of an answer before signing up. The public question feed now requires an account, so casual browsing stays private.

Big interface update

Howard got a full visual refresh and easier navigation on any screen size. New: ongoing conversations are grouped into threads you can pick back up later, your question history carries over, and guests get a real preview before signing up. Also new in the Tools menu: the Summary of Finances Downloader, which pulls SOF files straight from TEA.

Your feedback now goes somewhere

Added real feedback buttons, including a Great Answer button, so you can flag answers that are spot on or that missed the mark. What you flag goes to a review list that shapes what gets fixed next. Admins can also hide items from the public feed, and private questions stay private.

Better at finding the exact rule

A big upgrade to how Howard reads the rulebook. The account code tables in the back of FASRG are now indexed code by code, so questions like "what object code is this?" land on the exact row. A second-chance search kicks in when the first lookup is weak. Also added: the official documentation for TEA's EMAT ordering system.

A proper home: schoolbusinessbot.com

Howard moved from a test address to schoolbusinessbot.com. Old links still work, they just forward to the new address.

Howard tells you what it is doing while you wait

Instead of a blank pause, the chat bubble now shows progress while Howard searches and writes: looking up sources, reading the best matches, drafting the answer.

Accounts, sign in, and your own history

You can now create an account and sign in. Every question you ask is saved to a History page you can search later, so you never have to ask Howard the same thing twice. Admins got a simple panel for managing users.

The live feed

A new page shows questions as people ask them, in real time. It is a quick way to see what other school business folks are wondering about, and a reminder that you are not the only one with that question.

Howard learned ASCENDER

The complete ASCENDER business office documentation, more than 4,000 pages covering payroll, finance, budget, purchasing, personnel, asset management and warehouse, is now in Howard's library. Questions about where a screen lives or what a report shows now get real answers with links to the page.

Meet Howard

The school business assistant got a name, a face, and a new home at schoolbusinessbot.com. Same mission: instant, sourced answers to the school finance questions you'd normally dig through six PDFs to answer, with a link to the exact page so you can check the work.

Day one: the FASRG assistant

Before there was a Howard there was a plain chat box at fasrg.clapp.me, built in a single day. Today was all about getting the first version working and then immediately growing it.

• You could ask a question about FASRG v19 and get an answer with page citations. That was the whole product.
• Table-heavy pages like the revenue matrix came out as garbage at first. Fixed the same morning.
• By evening the library had grown past FASRG: all of EDGAR and 2 CFR 200 (the federal grant rules), plus the LASO program guidelines and FAQs. The name became the School Finance and Grants Assistant.
• Every question and answer started being logged so the School Finance Quiz could learn from what people actually ask.

Where it started

A zip file of FASRG v19 PDFs and an idea: years of school finance knowledge live in people's heads and in rulebooks nobody reads cover to cover. What if you could just ask?