UK chatbot beta launch: blog post, homepage CTA, featured tools

[vahid-ahmadi]

One PR consolidating the UK AI chatbot launch work (supersedes #1070 and #1072).

Blog post — "AI chatbot for policymaking" (beta)

• New post at app/src/data/posts/articles/uk-chat-cant-make-up-numbers.md + cover image, dated 2026-06-27.
• Framed as a beta launch; short noun-phrase section/subsection headings; six-tools figure (centred, grouped) and architecture diagram; numbered architecture/loop steps; key phrases bolded.
• Citations deep-linked to exact sentences (AI Playbook, Wikipedia, POST, NAO).
• posts.json: title "AI chatbot for policymaking", beta subtitle, date 27 Jun.

Homepage

• HeroCTA: second button "Try the AI chatbot" (navy, distinct from teal "Enter PolicyEngine") → PolicyEngine UK Chat.
• FeaturedResearchBanner: four cards refreshed to the latest UK tools, led by UK CliffWatch tool.
• HomeBlogPreview: excludes the chatbot launch post from "Expert policy analysis" (it's a product announcement, not policy analysis).

Apps

• Renamed UK CliffWatch → "UK CliffWatch tool" in apps.json and the home banner. (The CliffWatch apps.json entry itself already merged via Add UK CliffWatch to the apps registry #1071.)

Fixes

• MarkdownFormatter: restore ordered/unordered list markers that Tailwind preflight reset (numbered steps now show their numbers).

Note: the POST citation deep-link is best-effort — that page blocks automated fetching, so the exact-sentence highlight is unverified (falls back to page top if it doesn't match).

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
policyengine-app-v2	Ready	Preview	Aug 12, 2026 2:35pm
policyengine-calculator	Ready	Preview	Aug 12, 2026 2:35pm
policyengine-calculator-next	Ready	Preview	Aug 12, 2026 2:35pm
policyengine-website	Ready	Preview	Aug 12, 2026 2:35pm

[anth-volk]

Vahid, two requested changes in the “What you can ask” section:

• Please reword the introductory text to: “The Child Benefit reform above is one of four kinds of questions supported today, and users can stack these questions together within one conversation without starting a new chat.”
• In the dropdown from the “Kinds of Questions” visual, please add the label “Possible tool usage” before the boxes containing the tool names.

[vahid-ahmadi]

Vahid, two requested changes in the “What you can ask” section:

• Please reword the introductory text to: “The Child Benefit reform above is one of four kinds of questions supported today, and users can stack these questions together within one conversation without starting a new chat.”
• In the dropdown from the “Kinds of Questions” visual, please add the label “Possible tool usage” before the boxes containing the tool names.

done

[DTrim99]

Editorial review — grammar, clarity, house style

Reviewed against the beta-launch post ("AI chatbot for policymaking" / introducing-policyengine-uk-chat.md). No content questioned on substance — it's clean, well-structured, and appropriately neutral, and the PolicyBench framing ("the benchmark covers US policy, but the lesson carries directly to the UK") nicely pre-empts the obvious objection. Everything below is polish.

Corrections (grammar / house style)

1. "PolicyEngine is excited to announce…" (opening) — house style avoids "excited to announce" boilerplate. The branch already has a plainer variant ("PolicyEngine has launched PolicyEngine UK Chat"); prefer that, or "PolicyEngine UK Chat is now in beta." Dropping the enthusiasm also reads as more credible for a tool whose pitch is rigour.

2. The word "model" carries two meanings and collides. Throughout, "model" means both the language model (LLM) and the microsimulation model (PolicyEngine). Sharpest in the opening: "accuracy that models do not yet offer on their own: the figure has to come from a model whose assumptions can be inspected" — LLMs and PolicyEngine, one clause apart. Recommend a global convention: "language model" / "AI" for the LLM, "the engine" / "microsimulation" for PolicyEngine, and never bare "model" where either could be meant. Highest-value edit for comprehension.

3. "pairs the language strengths of large language models" — "language…language" redundancy plus a missing parallel "of." Suggest "pairs the linguistic strengths of large language models with the accuracy and testability of PolicyEngine UK."

4. UK spelling is consistent (modelled, programmes) — good; just confirm nothing slips to US spelling in later edits.

Clarity suggestions

5. "That scoped tool surface is the boundary…" — "tool surface" is engineering jargon a policy audience won't parse. Suggest "That limited, explicit set of tools is the boundary…"

6. "a reference generated from the installed engine — its capabilities and parameter schema" — strong and concrete but dense on first read. Consider "a machine-readable description of what the deployed engine can compute (its capabilities and parameter schema)". Keep the load-bearing point that it's tied to the deployed version rather than a drifting prompt — the best sentence in the section.

7. Link consistency — the final copy links the chat as [policyengine.org/uk/chat](/uk/chat), but an earlier draft in this branch points to policyengine-uk-chat.vercel.app. Pick one canonical URL and use it in the post, the HeroCTA button, and posts.json.

8. "matched PolicyEngine exactly on 80.3% of its scored outputs" — good that the final copy dropped the editorializing "only" an earlier draft had. Worth confirming the 80.3% figure and the GPT-5.5 model name are still current against the live PolicyBench post before publishing — it's the post's one hard number and model names/scores age fast.

Query

9. Title alignment. posts.json uses "AI chatbot for policymaking"; the article opens "…announce PolicyEngine UK Chat"; the PR description references "Bringing AI to policymaking." Settle on one title and make the H1/subtitle/description echo it.

Question on the post setup

While reviewing I couldn't fetch a single clean version of the post, because the branch/diff looks unsettled — flagging in case it's unintentional:

• There appear to be two competing article files — introducing-policyengine-uk-chat.md ("AI chatbot for policymaking") and uk-chat-cant-make-up-numbers.md ("Bringing AI to policymaking…") — with different headings and framing. Is one meant to be deleted?
• On the PR head I don't see either article file in app/src/data/posts/articles/, and posts.json at head has no entry for the new post — so as it stands, would merging actually publish it? The diff also carries built Vite artifacts (a large index.html, hashed JS/CSS/SVG under app/public/assets/posts/…) and some duplicated paragraph fragments, which reads like a stale base / needs-a-cleanup branch.

Could you confirm the intended final state — one canonical .md, a matching posts.json entry, and no build artifacts in the diff? Happy to re-review the prose once it's pinned to the version that'll actually ship; items 1–3 are the must-dos on the text itself.

🤖 Generated with Claude Code

[vahid-ahmadi]

@DTrim99 thanks — and your last section is the right question to have asked. The branch genuinely was confusing to review, so let me pin down the intended final state first, then what I've changed.

What actually ships

There is no markdown article any more, and that is deliberate. The post is now a scrollytelling page, following the pattern of the encode multi-agent post. Its source lives in PolicyEngine/policyengine-uk-chat-interactive-article with its own CI and review history, and it is served from its own Vercel deployment; this PR only carries the route and the registry entry. So:

• introducing-policyengine-uk-chat.md and uk-chat-cant-make-up-numbers.md are both gone — the second was superseded long ago, the first when the scrollytelling page replaced it. Nothing in posts.json: the entry is in apps.json with displayWithResearch: true, which is how the encode post and the other interactives are listed.
• No build artifacts in the final diff. You saw them because I briefly committed the built page here and then reverted it in favour of proxying to the deployment, and I chose not to force-push a branch that already had review comments on it. The commit list shows the churn; the file diff nets out clean. The post's files in this PR are now just apps.json, appZoneRoutes.ts, the cover image, and the changelog.

Your edits, applied

#2, the "model" collision — your highest-value point, and you were right. The word carried both meanings, sometimes a clause apart: "the tools UK Chat exposes to the model" (the LLM) near "run through the model to produce taxes, benefits, and net income" and "reports what the model calculates" (both PolicyEngine). I adopted your convention throughout — "language model" for the LLM, "the engine" or "PolicyEngine UK model" for the microsimulation, and no bare "model" where either could be meant. All 27 uses are now qualified, including the boundary diagram's accessible description and the section heading, now "The language model proposes a plan".

#6, the engine reference. Now "a machine-readable description of what the deployed engine can compute: its capabilities and parameter schema", keeping the point you singled out — that it is tied to the deployed version rather than a drifting prompt.

#7, one canonical URL. The article's call to action now points at policyengine.org/uk/chat rather than the vercel.app host. HeroCTA already resolves NEXT_PUBLIC_CHAT_URL || "/uk/chat", so the two agree. Worth being explicit, since it is also @MaxGhenis's first point: that route does not exist on production yet — it ships with this PR. Until this merges, the article's CTA 404s.

Already resolved on the page you'll re-review: "tool surface" jargon (#5) is gone; "excited to announce" (#1) and "pairs the language strengths" (#3) are not in it; the GPT-5.5 name is dropped and the 80.3% figure was re-verified against the PolicyBench post (#8); UK spelling is clean (#4); and the title now matches across the page, index.html and apps.json (#9).

One I have not changed, so you can object to it knowingly: the subtitle reads "Introducing PolicyEngine's newest AI-powered tool to help users understand UK tax and benefit policy". That is the closest thing left to the promotional register you flagged in #1, and it is the author's deliberate wording for a launch.

Where to re-review

The live article — the exact build that will serve at /uk/introducing-policyengine-uk-chat — is at https://policyengine-uk-chat-interactive-ar.vercel.app/uk/introducing-policyengine-uk-chat. Prose changes go in that repo rather than here.

🤖 Generated with Claude Code

[DTrim99]

Went through it again — thanks for the clear breakdown, that fully answers the setup question.

Branch integrity — resolved ✅

The net diff now lands exactly where you said: 9 files (routes, the apps.json registry entry with displayWithResearch: true, the HeroCTA two-button refactor, cover image, changelog) — no orphan .md, no build artifacts. The commit churn is just the built-page-then-reverted history, which nets out clean; not force-pushing a branch already under review was the right call. The scrollytelling-in-a-separate-repo structure (mirroring the encode post) makes sense.

Editorial items — 8/9 landed, and I verified them against the live article source

Re-checked each against src/App.jsx in the article repo (the live page is a client-rendered shell, so the source is authoritative):

• Build out Policies: empty state #2 (the "model" collision) — resolved, and this was the one that mattered most. All 27 model* uses are now qualified: the LLM is "language model" (24×), the microsim is "PolicyEngine UK model" / "the engine"; the heading is "The language model proposes a plan"; the boundary-diagram alt-text keeps them distinct. No bare ambiguous "model" left. 👍
• Create a component completion tracking dashboard #3, Create Policies: Create a policy with a reducer and log out the result #5, Duplicate app v1 GitHub Actions setup and launch with app v2 #6, Consider role of Vercel in app v2 #7, Launch basic site using Google Cloud App Engine #8, Integrate routing into template site #9 — all cleanly resolved. The redundant "pairs the language strengths" sentence is gone; no "tool surface"; the engine-reference sentence reads verbatim as suggested (the deployed-version-not-a-drifting-prompt point preserved); the CTA points at policyengine.org/uk/chat; "GPT-5.5" is dropped and the 80.3% figure stands alone and internally consistent; and the title is identical across the H1, index.html, and apps.json.
• Build out Policies: View my past policies #4 (UK spelling) — clean (modelled, programme, catalogue, artefact, behavioural, finalising; zero US spellings).
• Identify series of best practices that app will employ #1 — noted as your knowing decline, and I agree it's fine. No "excited to announce"; the kept subtitle ("Introducing PolicyEngine's newest AI-powered tool…") is light launch framing — descriptive, makes no policy claim, doesn't breach neutrality. Acceptable. If you ever want a hair cleaner, "newest" is the one word that adds nothing, but that's taste, not a fix.

Also confirmed anth-volk's "one of four kinds of questions supported today" reword is present in the "What you can ask" section.

Fresh pass on the new scrollytelling prose — no blockers, a few optional nits

Since the prose now lives in policyengine-uk-chat-interactive-article, these are for that repo, not this PR — all optional tightening:

1. Segment label: "AI model" vs "language model". The flow segments are tagged "AI model" while the body everywhere calls the LLM "language model" — two names for one thing. Unifying on "language model" would keep it airtight (it's not ambiguous, just inconsistent).
2. "stack these questions together" — drop "together" (redundant with "stack").
3. Two long/loose sentences worth a light split: the "…which is why the system has three parts rather than two…" gateway sentence, and the verbless Households caption "One household containing one benefit unit per call; more complex arrangements need separate calls."
4. Minor cohesion: "…because models can read loosely worded questions" → "because these models…"; and "the deterministic PolicyEngine engine" in the alt-text reads slightly doubled ("the deterministic engine" is cleaner).

The accessible descriptions on both diagrams are genuinely good (they describe the meaning and trace the flow, not just "a diagram of…"), the tool math is consistent (9+2+2+7+1 = 21), and punctuation/dashes are clean.

Net

The one substantive blocker from last time — that merging might not actually publish the post — is gone, and the prose is clean and publication-ready. Only open dependency I'd restate is the one you and @MaxGhenis already flagged: /uk/chat and /uk/introducing-policyengine-uk-chat don't exist on production until this merges, so the CTA 404s until then — expected for a launch PR, just worth a merge-day sanity check. Nothing here blocks from my side.

🤖 Generated with Claude Code