C:\Users\matt_\.claude\projects\C--Users-matt--OneDrive-Claude\memory\ and are indexed in MEMORY.md. Claude writes and updates these automatically. There are four types: project (ongoing work context), reference (where to find things), user (preferences and background), and feedback (how to behave).
Matthew moved in 2012 (earlier sessions had this as 2022 — this is the correct date).
- Andy Balding — technical lead — andy.balding@northernpowergrid.com
- Ben Hardy — complaints handler — ben.hardy@northernpowergrid.com
- Paul McGinlay — Operations Director — paul.mcginlay@northernpowergrid.com
- Louise Lowes — Director of Customer Service — louise.lowes@northernpowergrid.com (unverified)
- Garry Dando — Data Governance — garry.dando@northernpowergrid.com
New contacts: Mark Appleyard (Andy's manager) — mark.appleyard@northernpowergrid.com; Dean Hogarth (OCU DPO) — dpo@ocugroup.com; Heather Woodhead — General Manager Domestic Scale Connections, single point of contact since 9 September 2026 — heather.woodhead@northernpowergrid.com.
The crux (must stay central): Matthew could only have known about Job A by being told it on site in November 2024, because Job A appears on no document until NPG's 2026 works plan. He described it before that plan was disclosed, so his consent account is corroborated by NPG's own later document. Consent sequence settled 7 Oct 2026: nothing agreed or refused at the survey visit; the only consent ever given was on the doorstep on 2 November 2024, for the surveyed scope.
Status (17 October 2026): Woodhead sent a second works plan (verge excavation gone, Matthew's own supply now to be cut and rejointed) with works moved a fifth time to 8/9 February 2027. Matthew replied with a single request, produce the October 2024 survey record or confirm in writing it does not exist, with his June written-first position attached. Awaiting the Ombudsman case reference, then the supplementary package goes over.
Status (12 October 2026): Energy Ombudsman referral LODGED by email in two parts (four documents plus 28-exhibit bundle) and acknowledged same day. Key new evidence from SAR: NPG's own March 2026 Works Order answering "Wayleaves required? NO" and citing the 2024 complaint. Woodhead reply finalised and ready to send in-thread now the referral is a fact.
Status (6 October 2026): Complaint formally closed 29 July 2026 by final response signed Paul McGinlay. Energy Ombudsman window open to ~29 July 2027; referral not yet made. Full dispute mailbox harvested (80 emails) to _output\NPG-dispute\emails\. Decisive contradiction on record: McGinlay and Woodhead give conflicting official accounts of why the 2024 excavation stopped, neither evidenced. Woodhead letter 5 October refused any H&S investigation of 2024. Phase 1 works at No 20 scheduled 13 and 16 November 2026. Pre-November priorities: Ombudsman referral, HSE concern, reply to Woodhead, written answer on whether the November works affect No 18's supply. Full analysis in _output\NPG-dispute\notes\.
Do not use em dashes or hyphens as connectives in any written text (letters, documents, responses). Use full sentences instead.
Why: It makes writing look AI-generated. Matthew flagged this while reviewing NPG correspondence.
Exception: Hyphens in compound words (e.g. well-established, hand-pointed) are fine.
Matthew's instruction (5 October 2026): "Don't you send anything ever." Claude drafts emails; Matthew always sends them himself from his own client. There is no approved-then-send flow at all. No Outlook COM Send, SMTP, Graph, browser automation or any other transmission mechanism.
Why: Sending is outward-facing and irreversible; Matthew keeps that step entirely to himself.
Note: Memory file stored in the WYCA session memory store, but the rule applies everywhere.
On the WYCA interface, the Category column (ESOL / English / Maths / Digital / Construction) is custom to the pilot with no agreed WYCA definition. Learning Aim Title, SSA Tier 2 and Notional NVQ Level come straight from LARS and are solid; Category is derived by our own best-effort rules and is not validated.
Why: Chris Maunder worked out category manually from the ILR/DfE spec, not a lookup, and the dummy data assigns categories at random. The definition is WYCA's call.
How to apply: In any email to Chris/Pippa/WYCA about the interface, explicitly flag that Category is custom, undefined and best-effort, and ask them to confirm the definition.
mattnoodling@gmail.com — used by the Personal Assistant agent for outbound enquiries: quote requests, tradespeople, bookings, and similar personal tasks. Keeps Matthew's personal addresses out of agent correspondence.
Password stored in the global secrets vault as CLAUDES-GMAIL-PASSWORD.
C:\Users\matt_\.claude\sql-standards.md- Developer in all proc headers: Matthew Denison, Team: D&A
- UPDATE then INSERT — never use SQL MERGE statement
- RowHash computed in staging view, stored in target, compared on join
- Surrogate keys:
SK{TableName}ID— generated via@maxid + ROW_NUMBER()(no IDENTITY in Fabric) - DELETE commented out for Dims by default; active for Facts
- Schemas: target
dw, stagingSTG, procsLOAD
C:\Users\matt_\.claude\sql-standards.md- Fabric Warehouse throws
Incorrect syntax near the keyword '<alias>'for bare reserved word aliases - Always write
[RowCount],[Name],[Value]etc. — never unbracketed - Applies to ad-hoc queries and UNION ALL count scripts as well as stored procedure SQL
The disseminate.co.uk website is split into two repos in the diss-tfs Azure DevOps org, project Disseminate Web Site:
- disseminate-compass — Claude Compass internal guide site. Source:
_claude-guide\. Deployed tocompass.disseminate.co.ukvia Cloudflare Pages, triggered on push tomain. Password protected with Cloudflare Access (one-time PIN to matthew@disseminate.co.uk). - disseminate-web — Main disseminate.co.uk website (not yet built). Separate Cloudflare Pages project.
Hosting stack: 123-reg DNS → Cloudflare nameservers → Cloudflare Pages (free tier). Auth via Cloudflare Access (Zero Trust, personal plan).
Why separate repos: Each is an independent app — Compass updates should not trigger a main site deployment. The main site will be public; only Compass needs auth.
Claude Compass is styled after awwwards.com/jobs — bold, near-monochromatic editorial aesthetic with an amber accent.
- Sidebar:
#111111near-black - Accent:
#D4920Aamber/gold - Background:
#F2F2F2, cards white - Headings: 52px, weight 900, uppercase, tight letter-spacing
- Logo bars: amber shades
#F5D08A → #E8A020 → #D4920A - Shadows: minimal/flat. Border-radius: 6px.
- Registered name: Disseminate Ltd, company number 07705496
- Registered office: Suite 3b Kings House, 1 King Street, Leeds, LS1 2HH
- VAT number: 116365622 (standard 20%)
- Quote numbering: DIS-{year}-{nnn}, first issued DIS-2026-101 (Freedom CVR quote)
- Standard payment terms: invoice on completion, 30 days
~/.claude/CLAUDE.md is a symlink. The real file is C:\Users\matt_\OneDrive\Claude\_global\CLAUDE.md.
Always edit the real path directly — the Edit tool refuses to write through symlinks. The symlink keeps global rules in OneDrive for backup and portability while making them available in Claude's config folder.
When selecting from CRL_Silver_LH views, always use the exact column casing from the original LOAD proc (typically all lowercase, e.g. countryname, iso2). Alias to DW PascalCase in the #Staged SELECT. Everything referencing #Staged afterwards uses DW names.
Why: The Silver Lakehouse is case-sensitive. Using PascalCase like [CountryName] directly in a Silver query causes Msg 207 "Invalid column name" at runtime.
How to apply: Find the original Silver→STG LOAD proc and copy column names from its SELECT list exactly. Never infer casing from the DW or STG DDL.
Two CRL bugs shared one root cause: ROW_NUMBER() dedup with a non-unique ORDER BY. Fabric's engine breaks ties arbitrarily (results flip between runs) where legacy SQL Server happened to be stable, so latent bugs surfaced after migration.
01_load_crm_accountnotebook: tie on createdon picked a NULL industry row. Final fix (14/09): rank all candidate rows after joining to the account — CRM's own industry pointer first, then completeness, then recency; never dedup before applying the pointer. Check the deployed cell, not the repo copy, when debugging this notebookusp_DimDocumentDescription: partitioned by invoiceid while outputting ledgervoucher as DocumentID — invoice/credit-note pairs randomly dropped, phantom GL rows accumulated in the insert-only dim (~2,667 inflated fact rows). Runbook:CRL\_output\uat-fix-DimDocumentDescription-runbook.md
How to apply: partition by the column that becomes the output identity; make ORDER BY deterministic. Suspect this pattern in any migrated proc with unexplained row differences vs old CRL.
Known accepted issue: one duplicated fact transaction (Dream doc 27206 colliding with a D365 voucher); the Source guard in usp_FactFinanceTransaction is deliberately commented out. Same duplicate expected in PROD.
Once Matthew has approved a plan and execution is underway, proceed without giving additional warnings about actions covered in the plan.
Why: Rule 6 already covers this — plan approval = permission to execute. Mid-task warnings feel like re-asking for permission and slow things down.
How to apply: Give one-time warnings at the planning stage if something is irreversible or risky. Once Matthew says go, execute cleanly without commentary on each step.
When walking Matthew through a hands-on task, especially on a second machine he is operating while chatting (e.g. a client laptop), give exactly one route and one step at a time, then wait for the result before the next step.
Why: During the Savills laptop setup (Oct 2026) multiple alternative routes offered at once became overwhelming.
How to apply: Pick the single most likely route and commit to it. Give only the next action, in plain words, with the exact thing to type or click. Surface an alternative only after the chosen route has failed, and replace the plan rather than stacking options.
- Model: Jacuzzi J-235 (2021), nickname "Huzi"
- Purchased from: Outdoor Living — contact for servicing, parts, and warranty
- Tariff: Intelligent Octopus Go — 6.9p/kWh off-peak (23:30–05:30), 28.5p/kWh peak
- Setup: Built-in economy/off-peak mode + external timer plug
- Data retention: 2 years of transactional GL data only — do not load
bi.gl_groupedhistorical rollup (pre-2023 Dufrain-era monthly summaries with degraded dimensionality) - Invoiced filter: mainaccountid IN ('600000','600100','603500','619500','600500','601000','600600') AND postingtype <> 14
- AccDef filter: mainaccountid IN ('602000','602100','602200','602300','602010','602110','600000','600100','600500') AND postingtype = 14
- Sign convention: All amounts negated (* -1) to match business reporting convention
- Schema mapping:
_output\fee-flash-schema-mapping.xlsx - Source model: GL SL PBIP at
_context\powerbi\GL SL.SemanticModel\
- Symptom: a valid change lands via git sync, then Fabric silently overwrites it with the pre-sync version seconds later (autosave writes the stale in-memory copy back to git)
- Fix: close the notebook tab after running so Fabric drops its in-memory copy; the next sync then applies cleanly
- When Claude edits code and pushes: remind Matthew to close the tab in Fabric before "Update from Git", or immediately after his next run
- Do not suggest avoiding Fabric edits altogether — running notebooks in Fabric is unavoidable in Matthew's workflow
- Valid values: string, long, integer (not "int"), float, double, decimal, date, timestamp, boolean
- Symptom of mismatch: AssertionError "dataType None should be an instance of pyspark.sql.types.DataType"
- Root cause: getDataType() returns None when the string is not in its switch dict
- How to avoid: before adding a datatype value, grep nb_0001_functions_gold for getDataType or copy the datatype string from an existing column of the same type
- Applies to: tt-*.html, reapit-*.html, eufinance-*.html, and any future project docs across all customers
- Anchor slugs: Savills =
#savills, Turner & Townsend =#turner-townsend - Why: lands the user on their customer's block instead of the top of the page. projects.html is the canonical hub grouping all project docs
- Does not apply to: non-project pages (agent pages, home entertainment sub-pages, etc.) which may legitimately link back to their own parent
- Check
git remote -vbefore assuming platform tooling - Azure DevOps equivalent of
ghisaz repos, but Matthew prefers UI operations anyway - Direct PR URL pattern:
savills-uk.visualstudio.com/Savills%20Applications/_git/bi-data-platform-reapit-core/pullrequestcreate?sourceRef=<source>&targetRef=<target>
- Read-only commands (git status, git log, git diff, git show) are fine to use freely
- Exception: git operations directly implied by an agreed task instruction (e.g. "update this folder from branch X")
- Always ask or confirm before committing, pushing, pulling, or checking out
- Do NOT run
gh pr create,az repos pr create, or equivalents - Instead: write PR title, description, source/target branches, and a direct DevOps PR-create URL to
_temp\cmd-{project}-{timestamp}.txt - Open the file in Notepad so Matthew can copy-paste into the DevOps UI
AURORA_SAV_RPS_SAV; dev/QA/UAT use AURORA_SAU_RPS_SAU. Any SQL, config, or pipeline parameter targeting PRD must use the SAV name.- Discovered 4 Aug 2026: PRD initial load failed with "Schema does not exist or not authorized" — naming difference, not grants
- Fix: update
SchemaNamein the PRDCONFIG.Processesfrom SAU to SAV (PRD only — other environments keep SAU) - Fabric Lakehouse schema is
LANDINGin all environments — unaffected - PP schema naming in PRD is unverified — check before the PP PRD load
The load_bronze_deleted pipeline (cd_GetDeleted_Batch1 and cd_GetDeleted_Batch2) sources only from AURORA_SAU_RPS_SAU (49 table refs, 0 PP). The six PP tables in AURORA_PRP_RPS_PRP have no equivalent delete tracking.
- Confirmed 10 Aug 2026 while making the pipeline environment-aware (SAU→SAV in PRD)
- Not a blocker — Matthew flagged it as a gap to look at later
- A fix would mirror the SAU pattern: a PP base-vs-
_AFULL OUTER JOIN block set writing toLANDING.DELETED, or a separate PP deleted pipeline
eu-west-2.aws, database REAPIT_RAW.- Reapit SAU UAT —
ZL83042.eu-west-2.aws.snowflakecomputing.com(the account in the legacy Bronze notebooks) - Reapit SAU PRD —
ss94622.eu-west-2.aws.snowflakecomputing.com(carries the SAV-named schema) - REAPIT_PP —
sx94212.eu-west-2.aws.snowflakecomputing.com(separate property-management account) - Live host is held in a Fabric cloud connection, not the Variable Library. Credentials are not recorded here.
Silver layer is a Fabric Lakehouse (not a Warehouse). Views on the SQL analytics endpoint (EU_FO_Silver_LH.staging.*, EU_FO_Silver_LH.landing.*) require CREATE OR ALTER deployment scripts run against the Lakehouse endpoint.
- Update the SQL in the relevant notebook (
Create or Alter Staging Views.Notebook,Create or Alter Landing Views.Notebook) - Run the notebook against the correct workspace (dev / test / prod) to apply the change
- Do not try to
ALTERvia a Warehouse connection or Fabric UI — the mechanics differ
In EU_FO_Silver_LH, the landing.* objects are SQL endpoint views, invisible to Spark. The persisted-table notebooks (cte_*_persisted / usp_staging_*) read raw dbo tables unqualified and write delta tables to the lakehouse landing schema.
- Enum labels:
_$labelcolumns come from joiningdbo.GlobalOptionsetMetadataon entity, option set, and option value - Filters: landing views apply
recid IS NOT NULL; replicate in Spark - Placeholder dates: D365 uses
1900-01-01; null out before using in dedupe ordering, not after - .platform files: every Fabric item folder needs the hidden
.platformfile or git integration silently ignores the item - Endpoint sync lag: after a Spark write, the SQL endpoint can serve stale data until metadata sync — the
Sync SQL Endpointnotebook gates downstream pipelines; validate only after a refresh
September 2026: forecast payaways appeared twice in WIP reports (once as Payaway, once as Unknown). The warehouse held one clean row per transaction — the duplication was purely DAX: CALCULATE([x], 'Net Revenue Type'[col] = "value") overrides the visual's row filter on that column, so amounts reappear under other type labels.
- Fix — KEEPFILTERS around every Net Revenue Type predicate in
_Measures.tmdl: 6 measures on Is Revenue Or Payaway, 21 breakdown measures on Net Revenue Type - Gotcha — the Weighted measures returned 0.00 (zero weighting × amount) rather than blank, which still materialises phantom rows
- Status — applied manually in Desktop 15 Sep 2026; workspace repo copy to be re-pulled once merged to main
Audited 29 Sep 2026 by scripting every object out of every environment and diffing against main. Prod runs the As-Is versions with uncommitted in-place hotfixes, superseded by current main.
- Why — prod deployments are manual, and this change flips the Department column from local to central D365 name in reports, so it was held back
- Apply — check
eufinance-pbi243936-deployment.htmlbefore prod work on these dims; baseline must precede the master data enrichment; prod's DimDivision still has Title/DivisionTranslated (varchar(200)), DimDepartment still has DepartmentGroup
During the Master Data Fabric work (Aug 2026) Matthew repeatedly asked to slow down, and after a batch of unannounced edits: "im completely lost what did you just do. can you wait in future".
- Why — he follows along in the Fabric workspace and tests each change as it lands; batched changes break his mental model
- Apply — describe the change in plain terms before making it, wait for the go-ahead, confirm briefly after. Keep answers short and don't move on unprompted