Tenants at the LSE workshop made 28 asks, in five groups.
I turned them into 9 features. Eight came from the asks.
The scope also lists 13 things we will not build, each with a reason.
No accounts: they would store personal data, and the service is anonymous by design.
No live chat, said the scope: the service is labelled as AI, and points to Shelter and Citizens Advice. I built a voice version anyway on staging. It is the one no I took back.
Why it matters
A no without a reason looks like you did not listen. A written reason lets people check your thinking, and argue with it. It also stops the same ask coming back every month. And when you change your mind, the old reason is there for everyone to see.
Try this: write the reason next to every no. Then publish the list.
Workshop: LSE-led think tanks with tenants, summer 2026. The 13 come from the scope of works for the funded build, 2026. They are a separate list, not part of the 28. Voice experiment: staging, October 2026.
I ran the same model, 2.5 Flash, on the same 50 test cases.
On staging it scored 3.64 out of 5. On the live site it scored 2.66.
Staging has better retrieval and a better prompt. The model did not change.
So about one point came from the stack, not the model.
Why it matters
When answers are poor, the easy move is a bigger model. But a bigger model on a weak stack hides the real problem. It also costs more every month. Test the same model in both set-ups first. Then you know what the model is worth on its own.
Try this: before you pay for a bigger model, run your current one on your best retrieval and prompt.
Scores out of 5, 2 October 2026. Blind review by five Claude reviewers against the project’s legal context. One run per case. The reviewers are AI, not lawyers.
Choosing languages from census data, not national figures
National figures describe everyone, not your users.
The 2021 Census splits main language by housing, such as social renting.
212,725 social renters in England cannot speak English well or at all.
Their main languages set the list. The funded build offers 16.
Why it matters
The languages most spoken across a country are not always the ones your users speak. Picking from national figures can leave out the people who need help most. Cutting the data down to your users first is cheap. It also ties each language to a real number you can defend.
Try this: filter the census to your users, by housing, area or age, before you pick languages.
ONS Census 2021, England. My analysis of main language by tenure. 16 languages on staging, October 2026.
From a canvas straight into ShotGrid: what the pipeline checks first
Artists pick project, sequence and shot from live ShotGrid lists. Nobody types a shot name.
The pipeline’s own checks run before anything is sent. They cover the name, file type and version, and that the shot and task exist.
Version numbers come from ShotGrid’s next free number, never from a person.
A refusal comes back as a checklist the artist can fix, not an error code.
The last step goes to the desktop tool the studio already trusts.
Why it matters
A send that fails late wastes the artist’s time, and the pipeline team’s. Running the pipeline’s own rules first means a file is right, or the artist knows exactly why not. Calling those rules, not copying them, means they cannot drift apart.
Try this: if your tool feeds a system with rules, call its checks before you send. Do not write your own copy.
Shipped at AcmeAI & FX. Direct send shipped 7 August 2026. Hand-off to the desktop tool shipped 28 September 2026.
Clarify-first has a cost: what 82 conversations showed
The live assistant asks questions before it gives advice. That is clarify-first.
In testing, a simulated tenant said “I don’t know any more details”.
Then it said “please just give me your best advice”.
In 21 of 82 conversations on the live site, the assistant never got past questions.
This is a finding about the design. It is not a score for answer quality.
Why it matters
Asking first can make advice fit the case. But real people often do not know the details. If the assistant cannot move on, they leave with nothing. Set a limit on questions. Then give the best advice you can with what you have.
Try this: test your assistant with a user who says “I don’t know”. Count how often it never answers.
Blind evaluation, 2 October 2026. 82 cases, one run each, up to three assistant turns. The simulated tenant cannot answer questions its scenario does not cover.
Can I count your visit? If you say yes, Google Analytics records the pages you read, your device and your rough location. Nothing runs until you choose. What it collects