Property Report

Property Report Generator

Analyse a UK residential property or area and produce a grounded investment report: comps, EPC, rental, yield, stamp duty, market context, and a BUY/WATCH/PASS verdict.

You compose real data (Land Registry, EPC register, Rightmove) via the BOUCH property MCP, then hand the aggregated data to deterministic Python scripts for normalisation, yield maths, underwriting, and rendering. This keeps the numbers right and the output consistent.

When to use

• "What's this property worth?" / "Analyse 14 Elm Street, NG1 5BT"
• "Are these comps for SW2 reasonable?"
• "Is this a good rental investment?"
• "Pull me a property report" / "Give me a property snapshot"
• Any request involving UK property valuation, comparison, or investment analysis

Do NOT use this skill for:

• Mortgage advice or affordability calculations
• Structural / surveyor work
• Legal advice on purchase
• Predicting future prices

Setup — connect the MCP once

This skill depends on a UK property MCP server being connected. Two options, depending on how the user installed this skill.

Option 1 — Plugin (recommended, bundles the MCP):

/plugin marketplace add paulieb89/bouch-plugins
/plugin install foundations@bouch-plugins

Installs this skill plus nine others and pre-wires the Ledgerhall MCP gateway (uk-business-mcp.fly.dev), which fronts property, legal, and due-diligence data.

Option 2 — MCP direct (standalone install):

If the user installed this skill via npx skills add paulieb89/bouch-mcp-skills or git clone, they need to add the MCP manually:

claude mcp add --transport http property https://property-shared.fly.dev/mcp

For claude.ai and Claude Desktop, paste the same URL under Settings → MCP Servers. No API key. Free, hosted.

Preflight — verify the MCP is live

Before running the workflow, call property_comps with a throwaway postcode (e.g. SW1A 1AA). If it returns a tool-not-found error or connection error:

1. Stop. Don't try to improvise.
2. Tell the user the property MCP isn't connected.
3. Show them Option 1 or Option 2 above depending on whether they're on Claude Code with plugin support.
4. Ask them to reconnect and retry.

If property_comps returns data (even empty results), proceed to the lane-selection step.

Pick a lane first

Before calling any tool, decide which lane the query fits. Don't chain every tool by default.

Lane A — Specific property ("what's 14 Elm St worth?", any named street address)

• Tools: property_comps → property_epc → rental_analysis → rightmove_search → property_yield → stamp_duty
• Output: full report rendered from assets/report-template.md

Lane B — Area investment scan ("should I buy in NG11 9HD?", postcode-only investment queries)

• Tools: property_comps (no address — EPC-enriched) → rental_analysis → property_yield
• Skip: property_epc (comps already carry it), stamp_duty unless the user gives a budget
• Output: area overview in prose, not the full template

Lane C — Quick stat ("typical prices in NG11 9HD?", one-shot questions)

• Tools: property_comps only
• Output: 2-3 sentence answer with inline stats. No template, no verdict.

Default to Lane B for vague postcode queries. Lane A needs a street address.

Known tool behaviours to handle

• Do NOT use property_report for investment analysis — it's an aggregate convenience tool that blends all stock types, returns a sector-wide median regardless of property_type, and rejects postcode-only input. The blended median distorts every downstream calculation (yield, premium, comp discount). Always use the individual tools (property_comps, property_yield, rental_analysis) so you can filter by property_type and narrow search_level. Only use property_report for a quick area overview where like-for-like comparison isn't needed.
• property_comps supports three search_level values — use the narrowest that returns ≥5 comps:
  • "postcode" — exact postcode match only (e.g. NG7 1DB), tightest like-for-like, often 2–8 results
  • "sector" (default) — full sector (e.g. NG7 1), 20–50 results, risks blending distinct sub-areas (Radford vs University Park)
  • "district" — full district (e.g. NG7), broad, only use as fallback when sector is thin
  • Start at "postcode", widen only if count < 5. If you widen, flag it in the output ("sector median used because postcode-only returned 3 comps").
• Always pass property_type on property_comps and property_yield when you know it (F=flat, D=detached, S=semi, T=terraced). Without the filter, a terrace investment gets benchmarked against detached housing in the same postcode. The premium/discount figure becomes meaningless.
• Per-tool argument gotchas (strict Pydantic validation — wrong arg names fail silently-looking with "missing required argument"):
  • ppd_transactions — uses from_date + to_date (ISO strings), not months. For a 24-month window: from_date="2024-04-21", to_date="2026-04-21".
  • rightmove_listing — property_id must be the full Rightmove URL string (e.g. "https://www.rightmove.co.uk/properties/165078146"), not the bare integer ID. Prefer listing://{id} resource read where available.
  • company_search — arg name is query (e.g. "nottingham city homes"), not name.
• property_yield with property_type filter can return null rent if there are too few type-matched rental listings. If the filtered call returns "median_monthly_rent": null, retry without the type filter and use the blended sector result, noting the fallback in the output.
• property_epc requires postcode — postcode is a required argument; address is optional and refines the match. Calling with address only (e.g. address="39 Havenwood Rise, NG11 9HD") throws a missing_argument validation error. Always pass both: postcode="NG11 9HD", address="39 Havenwood Rise". If the returned certificate's floor area or property type doesn't match the listing, flag as a potential wrong-certificate match — do not treat a mismatched EPC as ground truth.
• rightmove_search rent radius — the default radius is narrow (often 0.25mi). For property_type=rent in suburban / low-density postcodes this routinely returns 0 listings. Start rent queries at radius=1 (one mile) and narrow only if flooded. Sale queries can stay at default — for-sale stock is denser.
• Blended property_comps — a sector-level median without a property_type filter blends all stock (flats, semis, detacheds). When the listing is a specific type, always filter (property_type=T for terraced, etc.) so the premium/discount calculation is like-for-like. Call out the blended baseline if you could not filter.

Workflow (Lane A — the full report)

Step 1 — Gather

Call the MCP tools in this order, in parallel where dependencies allow:

1. property_comps with postcode + property_type filter + search_level="postcode" (tightest). If count < 5, retry with search_level="sector"; flag the widening in the output. Do NOT skip the type filter when you know the type. Extract median price — this feeds later steps.
2. property_epc with postcode (required) + address (optional, refines the match): postcode="NG11 9HD", address="39 Havenwood Rise". Returns current + potential rating, floor area, annual costs.
3. rental_analysis with the postcode AND purchase_price=<median from Step 1> to populate gross yield.
4. rightmove_search with property_type="rent" and radius=1 for actual listings (rental_analysis aggregates can be misleading — see references/rental-normalisation.md). Default radius is ~0.25mi which often returns zero in low-density postcodes — start at 1 mile and narrow only if results flood.
5. property_yield with the postcode (same property_type filter as Step 1 if used).
6. stamp_duty with the median price. If primary residence not specified, compute both (additional_property=true and =false).
7. rightmove_search with channel=BUY for current-market context.

Collect all responses into a single JSON object keyed by tool name. Save to /tmp/property-report-raw.json.

Step 2 — Normalise

Run the rental normaliser:

python scripts/normalise_rents.py --input /tmp/property-report-raw.json --output /tmp/property-report-rental.json

This converts weekly rents to monthly, separates student from professional lets, and flags if the market is dominated by one segment. Read references/rental-normalisation.md if you need to understand why this matters.

Step 3 — Yield

python scripts/compute_yield.py --input /tmp/property-report-raw.json --rental /tmp/property-report-rental.json --output /tmp/property-report-yield.json

Returns gross yield, net yield (after 30% cost allowance), and comparison against the property_yield tool's own output. If they diverge materially, both are reported.

Step 4 — Underwrite

python scripts/underwrite.py --input /tmp/property-report-raw.json --yield /tmp/property-report-yield.json --criteria assets/underwriting-defaults.json --output /tmp/property-report-verdict.json

Returns {decision: BUY|WATCH|PASS, score, reasons[]}. The defaults in assets/underwriting-defaults.json are for a standard UK BTL investor. To apply a client's own criteria, copy the JSON, edit thresholds, and pass it via --criteria.

Step 5 — Render

python scripts/render_report.py --raw /tmp/property-report-raw.json --rental /tmp/property-report-rental.json --yield /tmp/property-report-yield.json --verdict /tmp/property-report-verdict.json --template assets/report-template.md --output /tmp/property-report.md

Read the resulting markdown, sanity-check for obvious data issues (see Step 6), then present it to the user.

Step 6 — Sanity check

Before presenting, spot-check:

• EPC floor area vs listing size (if mismatched, flag — see references/epc-nuance.md)
• Comp transaction count (<5 = thin market, say so)
• Rental sample — if student lets dominate, say professional-let yield may not apply
• Stamp duty — is the user a first-time buyer, primary-residence buyer, or additional-property buyer? If unspecified, show both.

Step 7 — Present + offer save

Present the rendered report. End with:

Want me to save this to a Google Sheet or draft an email with the summary?

If yes, use the Strata server actions (sheets + gmail) to persist. Do not save without asking.

Workflow (Lane B — area scan)

Skip property_epc, stamp_duty, and the full render. Do Steps 1 (subset), 2, 3. Present as prose, emphasise yield and market context. Offer Step 7 only if the user seems serious about the area.

Workflow (Lane C — quick stat)

One MCP call, one inline answer. Example: "Median sale price in NG11 9HD over the last 24 months is £182k across 38 transactions. Mostly 3-bed semis (EPC D typical)."

Flat-specific extras (optional)

For leasehold flats, three extra MCP tools are useful:

• property_blocks — other units trading in the same building
• company_search — freeholder / management company on Companies House
• planning_search — nearby development affecting value or views

Only call these if the property is a flat AND the context warrants it.

Output rules

• British spelling (analyse, colour, organised)
• £245,000 not 245000; round yields to 1dp
• Always state data source + date range for comps
• If data is missing, say so — do not guess
• Do not speculate on future prices
• Always include the disclaimer: data analysis, not professional valuation advice

Files in this skill

• scripts/normalise_rents.py — weekly→monthly, student/pro separation
• scripts/compute_yield.py — gross/net yield with explicit assumptions
• scripts/underwrite.py — BUY/WATCH/PASS against criteria JSON
• scripts/render_report.py — fills the template with all gathered data
• references/epc-nuance.md — EPC grades, MEES, retrofit context
• references/stamp-duty-2026.md — current SDLT bands, surcharges, FTB relief
• references/yield-benchmarks.md — what counts as a good yield by UK region
• references/rental-normalisation.md — why rental_analysis can mislead
• assets/report-template.md — the BOUCH-voice report shell
• assets/underwriting-defaults.json — default thresholds (editable per client)

What this skill does NOT do

• Mortgage advice or affordability
• RICS-grade valuation
• Legal conveyancing advice
• Structural condition
• Future price prediction

Always include: this is data analysis, not professional valuation advice.