TP Bennett Recent Projects Scraper

Purpose

Return tp bennett's recent architecture / interior-design projects, each enriched with who is involved (the named tp bennett project lead plus the client) and what is being done (sector, project stage/status, location, size, and a narrative summary). tpbennett.com is a Next.js site whose content is served from the Prismic headless CMS, which exposes an unauthenticated public JSON API. Prefer that API — it returns the identical content the site renders, structured, in one request, with no auth, cookies, proxies, or anti-bot stealth. Read-only; never writes.

When to Use

Monitoring tp bennett's latest completed and in-progress projects.
Building a feed of the practice's work with structured metadata (client, sector, location, status) rather than scraping rendered HTML.
Enriching a project list with the responsible tp bennett director/principal and a short description of the scheme.
Any flow that would otherwise scrape https://www.tpbennett.com/work/ — the Prismic API is faster, cheaper, and structurally reliable.

Workflow

The site's content lives in the Prismic repository tp-bennett. Its public Content API v2 is at https://tp-bennett.cdn.prismic.io/api/v2 — no auth, no cookies, no proxy, no stealth. Individual projects are documents of type work_item. Lead with the API; a browser fallback is documented at the end but costs ~100× more (the pages are fully JS-rendered).

Get the master ref. Refs rotate on every content publish, so fetch a fresh one each run:
```
GET https://tp-bennett.cdn.prismic.io/api/v2
```
Read refs[] and take the entry where isMasterRef === true (e.g. anxqLBIAAC0AuWDa).
List recent projects (ordered most-recent-first). URL-encode the query; brackets/quotes/spaces must be encoded:
```
GET https://tp-bennett.cdn.prismic.io/api/v2/documents/search
    ?ref=<REF>
    &q=[[at(document.type,"work_item")]]
    &orderings=[document.first_publication_date desc]
    &pageSize=100
```
Returns total_results_size (12 today, single page) and results[]. Use document.last_publication_date desc instead if you want "recently updated" rather than "recently added".

Optionally, the site's own curated recent strip (what the /work/ landing page shows) is the project_group document with uid=work-recent:
```
&q=[[at(document.type,"project_group")]]
```
Its data.projects[] is an ordered list of work_item content-relationships (currently 11-12-wellington-place, capital-group, dockside-canada-water). The embedded copy only carries name/title/type/colour/media — fetch the full work_item (step 3) for the enrichment fields.
Extract enrichment from each work_item (results[i].data):
- Identity / "what": name (project name), title (editorial headline), type (sector, e.g. WORKPLACE, COMMERCIAL, RETROFIT, LIVING), media.url (hero image).
- Structured facts: the slice with slice_type === "project_info" → primary.info[] is an array of { heading, text } rows. Observed headings: Client, Location, Size, Status, Sector, Discipline. Also primary.content[] (rich-text paragraphs — the project overview).
- "Who is involved": project_info.primary.quote_author — a named tp bennett lead with role, e.g. "Mark Davies, Director" / "Emma Green, Principal Director". Pair with the Client info row for the full "who". quote_author is null on some projects (e.g. masterplans) — treat as optional.
- Narrative / "what is being done": slices with slice_type === "header_paragraph" → primary.heading + primary.paragraph[] (each project has 1–2). These describe the scheme ("A new base for the company's main trading operations…").
- Flatten Prismic rich-text arrays by joining each block's .text.
Build the canonical page URL for each project: https://www.tpbennett.com/work/<uid>/ (e.g. https://www.tpbennett.com/work/bp-london/).
Emit the project list (see Expected Output). One page of results covers the full set today; paginate via &page=N only if total_pages > 1.

Browser fallback (only if the Prismic API is unavailable)

browse open https://www.tpbennett.com/work/ --remote
browse wait load then browse wait timeout 3000 — the grid is client-rendered and populates 1–3s after load.
Collect each project link (/work/<uid>/), then for each: browse open the detail page, wait load + wait timeout 3000, and browse get text body. The rendered detail page carries the same fields (client, location, size, status, discipline, the named lead, and the overview copy). A bare (no proxy / no stealth) session is sufficient.

Site-Specific Gotchas

The whole site is a Prismic-backed Next.js app. Repo id tp-bennett is discoverable from the homepage's inline Prismic script + images.prismic.io/static.cdn.prismic.io in the CSP. The API needs no auth, no cookies, no proxies, no --verified — a plain HTTP GET returns 200.
The ref expires. refs[0].ref changes on every content publish. Always fetch /api/v2 first and use the current master ref; a stale ref returns an error. Don't hardcode it.
Projects are work_item, not work. The work document type is the single /work/ index page (a bag of layout slices), not the project list. Querying document.type == "work" returns 1 doc with no projects.
project_group work-recent is the site's curated "recent" set (3 projects, hand-ordered). Querying all work_item ordered by first_publication_date desc gives the fuller recency-ordered list (12). Pick whichever "recent" you mean; this skill returns the full list and flags the curated subset.
"Who is involved" comes from within the project, not a team join. There is a project_team_member type (16 leadership/people docs), but its related_articles field does not reliably link members to specific projects (empty document links). Per-project attribution is project_info.quote_author (named tp bennett lead) + the Client info row. Don't waste time trying to join project_team_member → work_item.
quote_author can be null (e.g. masterplan/placemaking projects like al-jabal-al-aali, 33-canada-square). Emit lead: null rather than failing.
info[] headings are data, not a fixed schema. Read them dynamically (Client, Location, Size, Status, Sector, Discipline are what's observed today) — don't assume positional order.
type (sector) is free-text and sometimes compound — e.g. "COMMERCIAL, RETROFIT", "PLACEMAKING, LIFESTYLE". Casing is inconsistent (WORKPLACE vs Workplace vs Commercial). Normalize if you need canonical sectors.
Prismic image URLs have a doubled query string (...?auto=format,compress?auto=compress,format). It's harmless and served correctly — pass through as-is.
tpbennett.com 308-redirects to www.tpbennett.com; the API host tp-bennett.cdn.prismic.io does not redirect.
Pre-run anti-bot probe: none detected, and the successful API run required neither proxies nor a verified/stealth session. Only reach for the browser fallback (which does need wait load + wait timeout 3000 on the JS-rendered grid) if the Prismic API is ever down.
Browser extraction is ~100× the cost of the API for identical data — the autobrowse inner agent, given only the task description, independently chose the Prismic API over scraping.

Expected Output

{
  "success": true,
  "source": "prismic-api",
  "repository": "tp-bennett",
  "ref": "anxqLBIAAC0AuWDa",
  "count": 12,
  "curated_recent_uids": ["11-12-wellington-place", "capital-group", "dockside-canada-water"],
  "projects": [
    {
      "uid": "bp-london",
      "name": "BP London",
      "title": "Rethinking workplace inclusion",
      "sector": "WORKPLACE",
      "client": "BP",
      "location": "London, UK",
      "size": "253,000 sq ft",
      "status": "Complete",
      "discipline": "Interior Design",
      "lead": "Mark Davies, Director",
      "summary": "A new base for the company's main trading operations, spanning eight floors of the Cargo building in Canary Wharf; a fit-out designed around inclusivity and varied working styles.",
      "hero_image": "https://images.prismic.io/tp-bennett/aiBBSAeQX7-eWqnL_bp-hero-1-.jpg?auto=format,compress",
      "url": "https://www.tpbennett.com/work/bp-london/",
      "first_publication_date": "2026-06-03T15:05:06+0000",
      "last_publication_date": "2026-06-26T09:00:00+0000"
    },
    {
      "uid": "al-jabal-al-aali",
      "name": "Al Jabal Al Aali",
      "title": null,
      "sector": "PLACEMAKING, LIFESTYLE",
      "client": "Ministry of Housing and Urban Planning",
      "location": "Jabal Al Akdar, Oman",
      "size": null,
      "status": "Detailed Masterplan",
      "discipline": null,
      "lead": null,
      "summary": "...",
      "hero_image": "https://images.prismic.io/tp-bennett/...jpg",
      "url": "https://www.tpbennett.com/work/al-jabal-al-aali/",
      "first_publication_date": "2026-06-03T...",
      "last_publication_date": "2026-06-...T..."
    }
  ]
}

Failure shape (e.g. Prismic API unreachable and browser fallback also blocked):

{ "success": false, "source": "prismic-api", "error_reasoning": "Could not fetch master ref from https://tp-bennett.cdn.prismic.io/api/v2 (HTTP 5xx)." }

project-scraper