[PATTERN carta-writing-style v0.0.2] [PATTERN etiquette v0.0.6] [PATTERN text v0.0.8] [PATTERN tables v0.0.12] [PATTERN carta-watermark v0.0.10] [PATTERN base v0.1.0]

Budget scenarios

What-if modelling on top of an existing budget. Five scenario references grouped into trim and growth, plus a shared helper:

Trim:

• references/headcount-reduction.md — reduce headcount by a target %.
• references/revenue-shock.md — apply a haircut to revenue.
• references/cost-rebalance.md — open-ended, user states a cash goal.

Growth:

• references/new-fund-raise.md — model fee revenue uplift from closing a new fund.
• references/expansion-hire.md — add N new FTEs at a stated comp band.

Shared helper (used by Step 4):

• references/get-actuals.md — canonical YTD-actuals query, so scenarios are grounded in real spend rather than just budget assumptions.

Growth references can stack with each other when the user mentions multiple levers in one prompt (e.g. "raise a $500M fund AND hire 5 FTEs") — see the "Stacking" section in each growth reference.

UX Rules

Audience is an accountant in Excel. Plain English only. Never surface MCP identifiers, DWH column names, UUIDs, raw JSON, SQL, or step labels. Currency: $X,XXX positive, ($X,XXX) negative, totals bolded. Differences are absolute. Status: ✅ Match | ⚠ Mismatch ($X diff) | ❌ Missing in Carta | ❌ Missing in Client Doc.

Closing summary link is a workbook citation (<citation:Sheet!Range>) in Claude for Excel mode, and a file:// path in Claude Code / Cowork mode.

Every numbered choice in this skill — including the closing next-step menu — MUST be presented via AskUserQuestion. Never render options as a bare code-fenced markdown list. The AskUserQuestion tool is in allowed-tools; use it. Bare-text menus break the chooser UI in Claude for Excel and force the user to type the number.

When to use

Trim:

• "What if we cut headcount 10%?"
• "Model a 15% revenue shortfall"
• "Show me 3 trim options"
• "Propose 3 ways to preserve $500k of cash"
• "Compare two budget scenarios"
• "What happens if travel doubles?"

Growth:

• "What if we raise a new $500M fund?"
• "Model the P&L impact of Fund V closing in Q1"
• "Show revenue uplift from a $300M / $500M / $750M close"
• "Model hiring 5 FTEs in 2027"
• "Compare a 3 / 5 / 7 hire ramp"
• "How does next year's P&L look if we raise a new fund AND hire 5 FTEs?"

DO NOT use this skill for

• Building a new budget — use carta-create-budget.
• Refreshing actuals on an existing budget — use carta-fetch-actuals.
• Pacing / variance / "how are we doing" — use carta-budget-analysis.
• Historical / consolidating P&L or income statements (a P&L statement for a closed period, not a forward what-if) — use carta-consolidating-pnl. A forward "how would X affect our P&L" what-if belongs here.

---

Execution discipline

Execute all steps silently. Do not narrate tool calls, intermediate results, or status updates. Only speak at explicit decision points: Step 0.5 (if runtime is ambiguous), Step 1 (destination chooser), Step 2 (intent routing), Step 3 (parameter gate), Step 6 (approval), and Step 8 (next-step menu).

---

Entry mode — fresh session vs. chained skill

Mandatory telemetry beacon — run this before any skip-gate check: If <SERVER> and <ENTITY_UUID> are already in context, call mcp__<SERVER>__set_context(firm_id=<ENTITY_UUID>, _instrumentation={"plugin": "carta-investors", "skills": ["carta-budget-scenarios"]}) as your very first action. Do not skip this even when all data is already available — it records this skill invocation. If either is unknown, it fires in Step 0 below.

Before Step 0, check whether these context variables are already set from an earlier budgeting skill call in the same session:

• <SERVER> — connected Carta MCP server prefix
• <ENTITY_NAME> and <ENTITY_UUID> — the resolved entity
• <RUNTIME> — excel-addin or local-file

If all four are in context: skip Steps 0 and 0.5 entirely. Call mcp__<SERVER>__set_context(firm_id=<ENTITY_UUID>, _instrumentation={"plugin": "carta-investors", "skills": ["carta-budget-scenarios"]}) to re-anchor the session scope and record this skill invocation. Proceed from Step 1 (destination chooser).

If any is missing (fresh session or cold invocation): run Steps 0 and 0.5 in order, then continue from Step 1.

Do not ask "which firm?" or "which runtime?" when those are already established from the skill the user just ran.

---

Step 0 — Carta MCP environment + resolve firm

1. Call refresh_mcp_connectors. Filter servers[] to name matching Carta / Carta (…) / carta with status: "connected". Drop failed.
2. Probe all three prefix forms in parallel: mcp__claude_ai_Carta__welcome(_instrumentation={"plugin": "carta-investors", "skills": ["carta-budget-scenarios"]}), mcp__carta_production__welcome(_instrumentation={"plugin": "carta-investors", "skills": ["carta-budget-scenarios"]}), and mcp__carta__welcome(_instrumentation={"plugin": "carta-investors", "skills": ["carta-budget-scenarios"]}). First success = <SERVER>.
3. Don't call any other mcp__<SERVER>__* tool before welcome — every command is gated.

If none connected, list failed connectors and stop. If multiple, default to Carta (production).

Resolve firm: if user named one → mcp__<SERVER>__list_contexts(firm_name="<entity>", _instrumentation={"plugin": "carta-investors", "skills": ["carta-budget-scenarios"]}) → disambiguate via AskUserQuestion if multiple → mcp__<SERVER>__set_context(firm_id=<FIRM_UUID>, _instrumentation={"plugin": "carta-investors", "skills": ["carta-budget-scenarios"]}). Do not use call_tool for list_contexts or set_context — call the granular tools directly with _instrumentation as shown.

DWH param-name traps: dwh:execute:query takes sql: not query:. dwh:get:table_schema takes table_name: not table:. format accepts "ndjson" / "markdown", not "csv".

If no firm was named, defer to Step 3.

Step 0.5 — Detect runtime

Set <RUNTIME> to excel-addin (open workbook) or local-file (user-supplied path). If unclear, ask via AskUserQuestion.

Step 1 — Where should the scenarios live

Branches by <RUNTIME>.

If <RUNTIME> is excel-addin:

Empty-workbook shortcut: if the active workbook has one sheet, maxRows == 0, no other tabs (typically a fresh Book1.xlsx/Sheet1), skip the chooser. Announce the rename in one sentence — "I'll use the empty workbook you have open and rename Sheet1 to Scenarios." — then proceed. The chooser only exists to protect non-empty state.

Where should the scenarios live?

• "Add scenario columns next to the existing budget" (recommended for ≤3 scenarios).
• "Create a new Scenarios tab" (recommended for >3 scenarios or wide pivots).
• "Clone the workbook — leave the original untouched".

If <RUNTIME> is local-file:

Where should the scenarios live?

• "Add a Scenarios sheet to the same file" (recommended).
• "Write a separate <budget>-scenarios.xlsx file alongside the original" (preserves the original).

Store <DESTINATION> (open workbook + tab in add-in mode, or .xlsx path + sheet name in local-file mode).

Step 2 — Intent routing

Phrase	Reference
"cut headcount", "reduce salaries", "team reduction", "trim staffing", "headcount X%"	headcount-reduction.md
"revenue shortfall", "revenue haircut", "if revenue drops", "demand shock"	revenue-shock.md
"preserve $X cash", "hit a cash target", "free up cash", "propose ways to reduce spend"	cost-rebalance.md
"raise a new fund", "Fund <N> closing", "new fund raise", "AUM uplift", "management fee impact of a new fund"	new-fund-raise.md
"hire N FTEs", "add headcount", "expand the team", "hire ramp", "model new hires"	expansion-hire.md

Multi-lever prompts. If the user names more than one of the above in a single prompt (e.g. "raise a $500M fund AND hire 5 FTEs"), route to every matching reference. Combine their outputs into composed scenarios — each scenario column reflects the joint effect of all selected levers, and the cash-impact summary breaks out each leg (e.g. New-Fund Fees, New Personnel, Net). Do NOT run the references serially as separate outputs.

Immediately call read_skill for every matched reference — do not reconstruct scenario logic from memory:

Reference matched	Call
headcount-reduction	read_skill(file_path="references/headcount-reduction.md")
revenue-shock	read_skill(file_path="references/revenue-shock.md")
cost-rebalance	read_skill(file_path="references/cost-rebalance.md")
new-fund-raise	read_skill(file_path="references/new-fund-raise.md")
expansion-hire	read_skill(file_path="references/expansion-hire.md")

Step 3 — Parameter gate (batched)

Reference-specific. Generally include:

• Number of scenarios (default 3).
• Target % or target dollar amount.
• Scope — all lines vs a section (e.g. only Personnel, only Operating Expenses).
• Distribution rule when applicable (across-the-board, junior-heavy, senior-heavy, etc.).

Step 4 — Read the base budget + YTD actuals

If <RUNTIME> is excel-addin: use the add-in's read tools.

If <RUNTIME> is local-file:

uv run "${CLAUDE_PLUGIN_ROOT}/scripts/read_workbook.py" \
  "<BUDGET_PATH>" --sheet "<BUDGET_SHEET>"

Pull YTD actuals via references/get-actuals.md so scenarios are grounded in real spend rather than just budget assumptions.

Step 5 — Generate scenarios

Each reference computes its own scenarios. Outputs are always relative to the base budget as live formulas where possible — so the user can edit inputs and see scenarios recalculate.

• Trim references scale existing lines: =Base!H12 * 0.9 for a 10% trim.
• Growth references add new rows whose scenario columns reference user-editable named inputs at the top of the Scenarios tab: e.g. =fund_size * fee_rate * months_after_close / 12. Never hardcode the fund size, fee rate, hire count, or comp band — those must be editable cells so the user can flex them post-build.

Step 6 — Pre-build review (approval gate)

Preview table:

Section	Line Item	Base	Scenario 1	Scenario 2	Scenario 3	Recommended Δ
Personnel	Salaries	$4,200,000	$3,780,000	$3,360,000	$2,940,000	($420,000)

(Sample row shows the format — currency with $, negatives in parentheses, totals would be bolded. Column header is Recommended Δ regardless of trim vs growth — the value is the absolute change for the recommended scenario.)

Plus a cash-impact summary block at the bottom of the preview. Columns depend on the reference(s) that ran:

• Trim references: Scenario | Annual Spend Δ | Projected Cash at Year-End | NOI Δ
• Growth references: see the cash-impact-summary table in each growth reference (the columns name the specific lever — New-Fund Fees Y1, New Personnel Y1, etc.).
• Multi-lever (stacked) prompts: include one column per leg PLUS a Net column showing the combined effect.

Output the preview table and cash-impact summary above as a normal conversation message. Then call AskUserQuestion immediately after — the question field must be a single short sentence; never include preview content inside it.

• question: "Approve writing these scenarios?"
• header: "Approval"
• multiSelect: false
• options:
  1. Approve and write the scenarios ← recommended (description: "Writes the scenario columns to the destination chosen in Step 1.")
  2. Edit — change the target %, scenario count, or scope
  3. Cancel

The ← recommended marker goes inside the description field of option 1, not as a suffix on the label.

Wait for explicit OK.

Hard rule: no workbook-write tool (Excel-add-in cell write, execute_office_js that mutates state, write_workbook.py, or any equivalent) runs before this gate's AskUserQuestion returns the user's explicit "Approve and write" choice. If you catch yourself about to call a workbook-write tool without that approval recorded, stop and run this gate first.

Step 7 — Write and brand the tabs

Approval-recorded check (run FIRST, before any write tool)

Before calling execute_office_js with state-mutating code, setValues, write_workbook.py, or any other workbook-write tool, look back at your tool history. Find the most recent AskUserQuestion you sent. Does its answer literally include "Approve and write the scenarios"? If NO, Step 6 did not pass — send the Step 6 approval menu now and wait for the explicit answer.

Do not interpret upstream answers as approval. A parameter response, a routing answer, or any prior AskUserQuestion whose answer is not literally "Approve and write the scenarios" does NOT clear this gate.

Step 7 requires AT LEAST three separate execute_office_js calls (excel-addin runtime)

The most common failure mode is bundling cell writes + formatting + logo into one writeSheet(...) function — the model writes the cells, returns, and forgets the logo. Do not combine the cell-write call with the brand block in a single office.js block.

• Call 1: cell values, formulas, formatting. One execute_office_js. Return.
• Call 1.5 (blob pre-flight, REQUIRED before logo): verify the asset store is populated:

  const keys = blobs.keys();
  return { blobKeys: keys };
  

  If blobKeys is empty ([]), stop immediately — do NOT attempt the brand block and do NOT proceed to the summary. Surface this message to the user:

  "The Carta logo asset isn't loaded yet in this session. Please close the skill panel, wait a moment, then reopen it and try again. The scenario data has been written — only the branding step is pending." Do not silently skip branding. Do not mark the branding step as complete. Do not write Step 8 summary text.

• Call 2 (per tab touched): logo via the verbatim brand block from branding-and-header.md. Only run this after Call 1.5 confirms blobKeys is non-empty.
• Call N (verification, LAST): load shape names on every tab touched, confirm CartaLogo exists.

Returning from Call 1 does NOT finish Step 7. The verification call must appear in your tool history before Step 8 summary.

Before any write, call both of these in the same message (parallel reads):

1. read_skill(file_path="references/branding-and-header.md") — 4-row metadata band, logo placement, blobs.getText asset pattern, cell-comment API.
2. read_skill(file_path="references/<scenario-reference-from-step-2>.md") — the scenario file(s) matched in Step 2 (call one per reference for multi-lever prompts).

Do not reconstruct either spec from memory. All files must be in your context before generating any execute_office_js or write_workbook.py code. The branding-and-header.md file defines the reserved 4-row metadata band (A1 firm / A2 descriptive title like "2026 Budget · Senior-heavy trim scenario" / A3 source / A4 other context), the Carta logo placement (column E, rows 1–3 height), the blobs.getText("assets/...") asset-loading pattern for Excel add-in (NOT Read), and the cell-comment pattern for any low-confidence flag. If the existing Budget tab does not already have the 4-row band, add it as part of this write (shift data via sheet.getRange("1:5").insert(...)).

If <RUNTIME> is excel-addin: use the add-in's cell-write tools. Either side-by-side columns next to the existing budget, or a new Scenarios tab with a header section explaining each scenario's assumptions. After cell writes, brand every tab the skill touched — both the existing Budget tab (if a header band was inserted) and any new Scenarios tab — using the verbatim brand block in branding-and-header.md.

If <RUNTIME> is local-file:

uv run "${CLAUDE_PLUGIN_ROOT}/scripts/write_workbook.py" --stdin <<'JSON'
{
  "workbook_path": "<DESTINATION>",
  "operations": [ ... ]
}
JSON

Mark one scenario ← recommended in the cash-impact summary based on whichever best meets the user's goal (least painful trim that hits the target, smoothest NOI impact, etc.).

Branding verification (REQUIRED, observable, excel-addin only)

After running the brand block for every tab this skill touched, run this verification as a separate execute_office_js call before proceeding to Step 8:

const tabs = [/* "Scenarios", "Budget 2026", ... — substitute the actual tab names touched */];
const result = {};
for (const tabName of tabs) {
  const sheet = context.workbook.worksheets.getItem(tabName);
  sheet.shapes.load("items/name");
  await context.sync();
  result[tabName] = sheet.shapes.items.map(s => s.name);
}
return result;

The result must show CartaLogo in every tab's shape list. If any tab lacks CartaLogo, re-run the brand block for that tab and re-verify. Do not start Step 8 summary text until this verification returns CartaLogo on every tab.

Never silently skip branding. If the verification returns [] or an empty shape list, that means the brand block did not run successfully — do NOT proceed. If blobs were empty (Call 1.5 returned []), surface the reload message above and stop. If blobs were available but addImage still failed, surface the error and let the user decide.

Step 8 — Summary + next steps

Step 8 precondition (DO NOT SKIP). Before sending the summary text below, scan your tool history. Three anchors MUST be present in that order:

1. An AskUserQuestion whose answer included "Approve and write the scenarios" — Step 6 approval.
2. A sheet.shapes.addImage(base64) call for each tab the skill touched — Step 7 branding.
3. The branding-verification execute_office_js whose result showed CartaLogo on every tab — Step 7 verification.

If any anchor is missing, STOP, go back, and run the missing gate. Do NOT write "Carta logo placed at..." in the summary when no shapes.addImage call appears in your tool history.

Open the summary with a verb that matches the reference(s) that ran:

• Trim references → "Modeled 3 trim options…"
• new-fund-raise → "Modeled the fund-raise impact at 3 close sizes…"
• expansion-hire → "Modeled 3 hire-ramp scenarios…"
• Multi-lever (stacked) → "Modeled the combined impact of a new fund + N hires across 3 scenarios…"

If <RUNTIME> is excel-addin:

Modeled 3 trim options for Example MgmtCo. Scenario 2 (Senior-heavy reduction) preserves $487,000 of cash at year-end with the smallest impact on Q1 momentum — recommended. Full breakdown on Scenarios.

If <RUNTIME> is local-file:

Modeled 3 trim options for Example MgmtCo. Scenario 2 (Senior-heavy reduction) preserves $487,000 of cash at year-end — recommended. Full breakdown written to Scenarios in file:///path/to/<budget-workbook>.xlsx.

The next-step menu MUST be a single AskUserQuestion call with the options below as options entries. Never render them as a numbered markdown list, a bulleted list, or inline prose — bare-text menus break the chooser UI in Claude for Excel and force the user to type the number. The ← recommended marker goes inside the description field of one option, not as a suffix on the label.

1. Model a different scenario type (revenue shock, cost rebalance, new-fund raise, expansion hire) ← recommended
2. Drill into one of the scenarios — show me the impacted lines
3. Run a fresh pacing analysis using the recommended scenario as the new baseline
4. I'm done

Call AskUserQuestion with these exact parameters:

• question: "What would you like to do next?"
• header: "Next step"
• multiSelect: false
• options: the four label + description pairs above (place ← recommended in the description field of the recommended option, NOT in the label)

DO NOT render the menu as inline markdown text, a numbered list, a bulleted list, or closing prose. If your response is about to contain 1. ..., 2. ..., 3. ..., 4. ... as a list at the end of the summary instead of an AskUserQuestion tool call, you have failed this gate — back up and invoke the tool.

Mark option 1 ← recommended after the first scenario set; option 3 if the user is exploring "what would the new pacing look like".

When the user selects an option, immediately invoke the corresponding skill via Skill('<skill-name>') BEFORE doing any work. Do not freelance the output — load the downstream skill's SKILL.md so its gates, layout spec, branding rules, and approval flow apply. Routing:

Option	Skill to invoke
1 — Model a different scenario type	Skill('carta-investors:carta-budget-scenarios') re-entry with the new scenario type
2 — Drill into one of the scenarios	Stay in this skill — render the impacted-lines breakdown inline
3 — Run a fresh pacing analysis using the scenario as baseline	Skill('carta-investors:carta-budget-analysis')
4 — I'm done	No invocation; close cleanly

---

DWH result formatting

Queries > 50 rows: request format: "ndjson", bucket into a blob. Don't paste large results — triggers context_snip. Use "markdown" only for ≤50-row previews.

Hard rules

• All scenario values are live formulas referencing the base budget — never hardcoded duplicates.
• Recommended scenario needs a one-sentence rationale in the cash-impact summary.
• Local-file: openpyxl preserves formulas; scenarios use ='Budget 2026'!H12 * 0.9 syntax (single quotes around sheet names with spaces).
• Scenario labels are mechanistic, not sentiment-based. Use Across-the-board / Junior-heavy / $300M close / Base + Fund V. Never Bull / Base / Bear, Optimistic / Pessimistic, Best case / Worst case — describe what changed, not how to feel.
• Two-row header for month-bucketed tables. Row N = merged month label. Row N+1 = sub-headers. Never write both into the same row — subsequent merges destroy sub-headers.
• range.merge(true) discards trailing cells. Insert a new row first.
• Month-label date-serial trap: prefix with ' or use numberFormat: "mmm yyyy" on a real date.
• Label/header text beginning with +, =, -, or @ parses as a formula. Prefix such text with a leading ' (or a space), or drop the leading symbol — including Δ-delta and + Δ scenario column headers.
• Set Worksheet.position in a separate context.sync(), never in the same statement as worksheets.add(). Create and activate first: const sh = sheets.add(name); sh.activate(); await context.sync();.
• Currency — derive from the data, never default to USD. Resolve the workbook's presentation currency before writing (entity properties via welcome, or the currency on the base budget data); if it can't be resolved, ask the user. Use the locale-specific format token — never a bare $ (renders as system symbol on non-US locales): USD [$$-en-US]#,##0.00_);([$$-en-US]#,##0.00);"-", EUR [$€-x-euro2]#,##0.00_);([$€-x-euro2]#,##0.00);"-", GBP [$£-en-GB]#,##0.00_);([$£-en-GB]#,##0.00);"-", CAD [$CA$-en-CA]#,##0.00_);([$CA$-en-CA]#,##0.00);"-". The A4 band reads Amounts in <resolved_currency>.
• Border syntax (Office.js): style = "Continuous", then weight = "Thin". Never style: "Thin".
• Recalc + column widths (excel-addin): the last statements in the cell-write execute_office_js block (Call 1), in this order — never a separate call: restore automatic calc → context.workbook.application.calculate(Excel.CalculationType.full) → set fixed widths on label columns → autofit numeric columns only → context.sync(). Pattern:

  context.workbook.application.calculationMode = Excel.CalculationMode.automatic;
  context.workbook.application.calculate(Excel.CalculationType.full);
  sheet.getRange("A:A").format.columnWidth = 120;  // Section — fixed
  sheet.getRange("B:B").format.columnWidth = 180;  // Line Item — fixed
  sheet.getRange("C:<last_col>").format.autofitColumns();  // Base + scenario columns only
  await context.sync();
  

  Why fixed widths for label columns: autofitColumns() on the full range sizes column A to the widest section name, making it excessively wide. Always set A and B explicitly, autofit only C onwards. Recalc before autofit: scenario values are live formulas — without the forced recalc they stay at 0 and accounting format shows - (then #### after autofit). In local-file mode, add a set_column_width op for columns A and B, then autofit_columns over C onwards. Never autofit a header-only range.
• Branding standards — follow references/branding-and-header.md for every tab. Rows 1–4 reserved, logo at column E, blobs.getText("assets/...") for asset access.

---

Error handling

Never auto-retry. Always surface the failure and let the user decide.

• No Carta MCP connected → "Open Settings → Connectors, enable Carta, retry."
• Headcount-reduction: no personnel lines → echo closest matches (Salaries/Wages/Bonuses/Benefits), ask user which to include.
• Revenue-shock: no revenue rows → surface sections found, ask which lines are revenue.
• Cost-rebalance: no cash cell/tab → ask user for current cash.
• New-fund-raise: no Income section → echo closest section labels, ask where to add Mgmt Fee — <Fund>.
• New-fund-raise: non-numeric fund size → ask for the specific dollar figure(s).
• Expansion-hire: no Personnel section → fall back to headcount-reduction.md GL patterns, confirm placement.
• Expansion-hire: loadings unclear → default payroll tax 8% / benefits 20% / bonus 15%, show defaults so user can override.
• User uses sentiment labels (Bull/Base/Bear) → acknowledge, label scenarios mechanistically in output ($300M close, $500M close). Never carry sentiment into the workbook.
• Recommended scenario doesn't hit target → surface gap, ask whether to widen lever set.
• Auth error → ask user to reconnect Carta. Do not auto-retry.
• Connector connected, tool calls fail (McpAuthError) → prefix mismatch, NOT auth. Re-run refresh_mcp_connectors, probe matching prefix's welcome. Never tell user to re-auth without verifying.