data-cleanup

Data Cleanup

Two jobs in one skill for the recurring "my CRM is messy" pain (Sidekick's #1 negative-sentiment JTBD):

• Health check (report-only): a scored structural audit of a CRM board — missing-data patterns, stalled groups, abandoned columns, automation gaps, and a 0–100 data-quality score — published as a diagnosis doc. No writes.
• Cleanup (fix): audit the board for data gaps, let the user pick which to fix, then run a bounded, idempotent write loop.

Flow (report): Trigger → Gather → Score (5 strands) → Publish diagnosis doc → offer to fix. Flow (fix): Trigger → Audit → User picks gaps → Plan writes → Execute (bounded, idempotent) → Summary.

Input

• Optional: board name/ID via argument.
• Optional: pre-declared mode (Default / Silent / Proactive).

Output

• α (default): Bounded write loop on the chosen board. Normalizes phone/country/email; can also batch-set owner / source / last-touch / stage when the user explicitly picks those gap types and signals are clean. A Hygiene fix — <Mon DD> doc lists every change with Generated by Claude footer + <!-- claude-skill-id: data-cleanup -->.
• β (opt-in): A "before/after" snapshot doc (preserves the prior state for ~30 days, in case a pass over-corrected). User is asked at Step 4.
• Proactive extension: Out of scope — data-cleanup is already a write skill; proactive doesn't change its behavior. Default mode + batched confirm is the only write surface.

Knowledge

• Gap taxonomy + audit rules (§ Step 3).
• Normalization recipes (§ Step 5).
• Shared artifact conventions (§ Shared patterns).

Tools (MCP)

• get_user_context, list_workspaces, search, get_board_info, get_column_type_info, get_board_items_page, board_insights — Gather.
• list_users_and_teams — enumerate users for bulk-assign owner step.
• change_item_column_values — the core write. Loop with bounded concurrency.
• create_doc / create_update — publish summary + before/after snapshot.
• all_monday_api — escape hatch for column types not exposed by change_item_column_values. For country columns, use: change_multiple_column_values with column_values: {"<country_col_id>": {"countryCode": "<ISO2>", "countryName": "<name>"}}. Do not freeform the mutation type — only this shape is supported for country columns.

Cross-skill handoffs

• From daily-briefing: if the daily brief flags >N hygiene issues or >30% unowned deals, suggest this skill (health check first, then fix).
• From forecast / workspace-builder: if forecast-critical fields are missing or a freshly built board needs a fill-rate audit, route here.
• To daily-briefing / forecast: clean data feeds them — point the user there once cleanup is done.

Step 0: Connector check

Goal: Fail fast if the user has no monday MCP connection — no partial writes, no guessing.

1. Try mcp__monday__get_user_context.
2. If the tool is missing, returns an auth-style error, or the user has no accessible workspace: print "I don't see the monday connector active on this session. Install it from https://monday.com/mcp, then run this skill again." and stop.
3. If it works, cache user.id + user.name for artifact metadata.

Step 1: Detect mode

• Default — confirm before each batch (not each item) of writes.
• Silent — skip batch confirms; still publishes the summary doc.
• Proactive — same as Default for this skill (proactive doesn't make sense — every write is intentional). If user pre-declared Proactive, treat it as Default and note in chat.

Hard safety rail regardless of mode: no deletes, no amount-column writes, no cross-workspace moves, no normalization on columns the user didn't pick. Stage / owner / source / last-touch edits ARE allowed when the user explicitly selected that gap type at Step 4 — data-cleanup is the canonical place for board-wide column rewrites that other skills point users at.

Also detect report vs. fix intent:

• Report ("health check", "audit", "what's broken", "what are the data gaps", "board is messy") → run Step 2.5: Health-check report (report-only), publish the scored diagnosis doc, then offer to proceed to fixes.
• Fix ("clean", "fix", "normalize", "fill in") → skip to Step 3.
• Ambiguous → ask once: "Want a read-only health check (scored report, no changes) or to clean up the data now?"

Step 2: Resolve board

Goal: Land on a single boardId before reading any data.

1. If the user passed a board via argument: numeric → use as boardId directly; string → mcp__monday__search with searchType: "BOARD". One match → use it; multiple → AskUserQuestion.
2. If no argument: get_user_context → scan relevantBoards + favorites for names matching deals|opportunities|pipeline|leads|sales|contacts|accounts. One candidate → use it; multiple → AskUserQuestion; zero → list_workspaces → search. Still zero → stop with "I don't see a CRM-shaped board in your workspaces. Tell me the board name or ID, or run /monday-crm:workspace-builder to create one."

Step 2.5: Health-check report (report-only path)

Goal: Produce a scored structural audit as a doc — no writes. Run only when the user asked for a report/health check (see Step 1 intent). Otherwise skip to Step 3.

Pull board schema (get_board_info, get_column_type_info), items (get_board_items_page, up to 2K most-recent), aggregate stats (board_insights — known ~9% baseline error rate; on failure skip that strand and note it), and activity (get_board_activity, or derive staleness from item updated_at). Then score five strands:

1. Missing data — required fields (stage, owner, value, close_date) blank on >25% of items; compute fill-rate per column.
2. Stalled groups — groups with no item updates in >30 days.
3. Column drift — columns populated on <5% of items over 90 days (likely abandoned → suggest archive).
4. Automation gaps — stage-based automations absent on deal boards (probe via all_monday_api; if scope unavailable, skip with note).
5. Structure — duplicate columns, no groups, missing required column type.

Scoring (each strand 0–100, floor 0): Missing data 100 − (15·P0 + 8·P1 + 3·P2) · Stalled 100 − 15/group · Drift 100 − 8/column · Automations 100 − 20/missing stage automation · Structure 100 − 10/issue. Overall = weighted avg: Missing 40% + Stalled 20% + Drift 15% + Automations 15% + Structure 10%. Severity: P0 = fill-rate <50% on required fields, stall >90d, or no stage automation (blocks forecast/reporting). P1 = fill-rate 50–75%, stall 30–90d, column <5% populated. P2 = structural/minor.

Publish: create_doc titled Board Health Check — <board> — <Mon DD>, body carries Generated by Claude · <ISO> + <!-- claude-skill-id: data-cleanup -->. Sections: overall score + 5-strand breakdown, P0 fixes ([P0] <issue> → <fix>), P1/P2 batched, automation recommendations, column-archive list. Healthy-board shortcut: score ≥90 and zero P0/P1 → skip the doc, print the score + inline P2 notes.

Then offer to fix: "Want me to fix the fixable gaps now? I can normalize phone/email/country and bulk-set owner/source/stage within a scope you pick." → yes routes to Step 3; no stops here. Report path writes nothing.

Step 3: Audit — surface the gaps (Gather + Synthesize)

Goal: Compute a gap report grouped by fix-type, with a concrete count per type.

Pull all items via get_board_items_page (paginated up to 5000, then sample with disclosure). For each item, check:

"Company resolvable" definition for email guessing: look for (a) a text or email column containing a domain (@ or .), (b) a board_relation column linking to an accounts board, or (c) an item name matching First Last @ CompanyName. If none found, mark the item as not guessable.

Resolve columns by type, not English name (same rule as daily-briefing). Skip gap types where the matching column doesn't exist.

Produce the audit summary in chat:

## Audit — <board name>

**Items scanned:** N
**Gaps found:**
- ✅ Fixable (column normalization): <N> phone formats · <N> mixed-case emails · <N> missing country codes · <N> guessable missing emails
- ✅ Fixable (bulk-set within picked scope): <N> missing owners · <N> stale last-touch · <N> missing sources · <N> stage drift
- ⚠️ Flag-only: <N> missing phones (no source) · <N> missing amounts (forecast integrity rail)

Pick which fixable gaps to run, or "all". Flag-only items list in the summary doc — no writes.

Step 4: User picks (Synthesize)

First, ask whether to create a before/after snapshot: "Want a snapshot doc so you can see exactly what changed and roll back individual cells if something looks off? (yes / no)"

Then AskUserQuestion, multiSelect:

• "Phone formats (N items)"
• "Mixed-case emails (N items)"
• "Country codes from phone/email (N items)"
• "Guess missing emails from company domain (N items) — heuristic, may miss"
• "Bulk-assign owner (N items) — requires picking a scope and a target owner"
• "Backfill stale last-touch (N items) — will only fill empty or older-than-90d cells"
• "Bulk-set source (N items) — requires picking a scope and a source value"
• "Normalize stage drift (N items) — requires picking a scope and target stage"
• "All column-normalization fixes" (the first four — bulk-set ones still need scope)

Scope picker for bulk-set options: if the user picked multiple bulk-set options (owner / source / stage / last-touch), ask once: "Should the same scope (group / filter) apply to all of them, or do you want to set each separately?" One question beats four. If "same for all" → ask scope once and apply it across all bulk-set options. If "separately" → run a scope sub-prompt per option.

For bulk-set options, use list_users_and_teams to present available owners/teams when the target owner needs to be selected.

If user picks "Guess missing emails", surface a strong warning: "Email guessing uses a <first>.<last>@<companydomain> pattern. Accuracy varies widely — works best for B2B companies with standard email formats, unreliable for personal domains or non-Western name conventions. Only recommended for Leads boards where invalid sends are tolerable. Continue?"

Hard rule: no fix runs without an explicit Step 4 selection (and a scope-picker confirm for bulk-set options). Empty selection → stop with summary-only.

Step 5: Plan + execute (Publish)

Goal: Bounded, idempotent loop. Every write is reversible.

For each picked gap type, build a write plan: list of (itemId, columnId, oldValue, newValue, reason).

Normalization recipes

• Phone format → E.164:
  • Strip non-digit/+ characters.
  • If no leading + AND country column is populated → prepend dial code per country.
  • If no leading + AND country empty AND phone starts with a known country prefix (e.g., 1 for US, 44 for UK) → prepend +.
  • Edge case: phone is too short (<7 digits) or too long (>15 digits) → skip with reason "unparseable".
• Email lowercase: value.toLowerCase(). Only write if changed.
• Country code from phone: parse leading dial code → ISO-2 country code → write to country column using all_monday_api with {"countryCode": "<ISO2>", "countryName": "<name>"} shape.
• Country code from email CCTLD: if email ends in .uk / .de / .fr etc. and country empty → infer.
• Email guess: for items with name + resolvable company domain, try <firstname>.<lastname>@<domain> lowercase. Add a Source = guessed status flag on the item so the user can filter and verify later. If a source/status column is missing, prompt the user once to add one — never write the email without a verifiability flag.
• Bulk-set owner / source / stage: within the user-confirmed scope (group / status filter from Step 4), apply the chosen target value via change_item_column_values. Never extend beyond the picked scope.
• Backfill stale last-touch: set to today's date only on items where last-touch is (a) empty, or (b) older than 90 days AND user explicitly picked the "stale last-touch" gap type. Never overwrite a more-recent date.

Bounded execution

1. Cap total writes at 500 per run. Above that, prompt: " writes planned, capped at 500 per run. Run first 500 now and re-trigger for the rest, or narrow your selection?"
2. Execute via change_item_column_values, batched by item (one call per item updates multiple columns at once). Concurrency: 5 in flight max to avoid rate limits.
3. Confirm before write (Default mode) — single batched plan:
   • "Plan: phone normalizations · email lowercase · country fills · owner assignments (scope: ) · source backfills (scope: ) · stale-last-touch fills · stage drift fixes (scope: → ). Proceed? (yes / show 3 examples per fix-type / no)."
4. After each batch of 50 writes, print progress: Updated 50/500 cells. Continuing...

# Hygiene fix — <board name> · <Mon DD>

<!-- claude-skill-id: data-cleanup -->

## Writes executed (<N>)
| Item | Column | Old → New | Scope / reason |
|---|---|---|---|
| ... | ... | ... | ... |

## Skipped (<N>)
| Item | Column | Reason |
|---|---|---|

## Flag-only (no writes)
- N items missing phone (no source — manual entry required)
- N items missing amount (forecast integrity rail — review manually)

## Undo
Every write here is reversible. The "Old" column above shows the prior value. To revert: find the row in monday and copy the Old value back in manually.

---
Generated by Claude · <ISO timestamp>

Board Diagnosis — Sales Pipeline — Jun 15

<!-- claude-skill-id: data-cleanup -->

**Score: 62 / 100**

| Strand | Score | Detail |
|---|---|---|
| Missing data | 55 | 12 deals missing Owner (fill-rate 78%) |
| Stalled groups | 40 | 8 deals stalled >30d in "Discovery" |
| Column drift | 85 | "Legacy Status" column — 0% fill, candidate for archive |
| Automations | 60 | No stage-move automations on this board |
| Structure | 90 | Board structure looks sound |

**P0 fixes (2)**
- [P0] Owner blank on 12 active deals → batch-assign in bulk-data-hygiene
- [P0] 8 deals stalled in "Discovery" >30d → review + advance or lose

**P1 fixes (1)**
- [P1] "Legacy Status" column — 0% fill over 90d → archive via Board Settings

**Automation recommendations**
- Add status-change → email-owner automation for stage moves

---
Generated by Claude · 2026-06-15T09:00:00Z

# Hygiene fix — Sales Pipeline · Jun 15

<!-- claude-skill-id: data-cleanup -->

## Writes executed (3)

| Item | Column | Old → New | Scope / reason |
|---|---|---|---|
| Acme Corp — Q3 expansion | Phone | +1-555-123-4567 | +15551234567 | E.164 normalize |
| Beta Industries — pilot | Email | [email protected] | [email protected] | lowercase |
| Gamma Ltd — renewal | Country | (empty) | GB (United Kingdom) | from email CCTLD .co.uk |

## Skipped (0)

## Flag-only (no writes)
- 12 items missing Owner (no source — select scope then re-run to bulk-assign)
- 2 items missing Amount (forecast integrity rail — review manually)

## Undo
Every write here is reversible. The "Old" column above shows the prior value.

---
Generated by Claude · 2026-06-15T09:00:00Z

Cleaned 3 cells across 3 items on Sales Pipeline. 0 skipped. Doc: [url].