Carta Consolidating Financial Reports

[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]

Consolidating Financial Reports

Single entry point for the firm-wide, multi-entity financial statements Carta builds in Excel. Routes to one of three reports, or all three together:

• Consolidating P&L — income statement across the entities you pick, with a detail tab and an executive Summary tab. Optional tag/department view.
• Consolidating balance sheet — Assets / Liabilities / Equity side by side by entity, with a Total column.
• Consolidating trial balance — a flat consolidated tab plus a by-entity tab.
• All three together — builds all three in one pass, resolving the firm, entity scope, period, and destination once instead of three times.

Every report reads live from Carta Fund Admin and writes into Excel. The audience is an accountant working in a workbook, not an engineer.

All three reports are implemented inline, from this skill's own references/ tree. There is no dispatch to a separate skill — this skill owns the whole user-facing surface for the consolidating-financials theme.

Route The Request

Use this table to jump straight to the right report. Determining the report (this table, or the Router Gate's AskUserQuestion menu) happens before Gate 0 and Gate 0.5, so the user sees what this skill can do without waiting on any MCP round-trip. Actually loading the report reference, however, waits until Gate 0 has resolved the firm, the entity scope, and the period, and Gate 0.5 has established the surface and the target workbook — never load a report body before both gates pass.

If the user's prompt matches multiple rows or is ambiguous, fall through to the Router Gate — that section carries the AskUserQuestion disambiguation menu.

Customer Intent Framework

Use this as the semantic layer when the Router Gate's phrase table doesn't produce an exact match — before falling back to the welcome screen + AskUserQuestion menu.

UX Rules

Audience is an accountant in Excel. Plain English only. Never surface MCP identifiers, DWH column names (ACCOUNT_TYPE, EFFECTIVE_DATE), UUIDs, raw JSON, SQL, or gate labels.

• Every numbered choice in this skill — including the report menu and the entity picker — MUST be presented via AskUserQuestion. Never render options as a bare code-fenced markdown list. Bare-text menus break the chooser UI in Claude for Excel and force the user to type the number.
• The Router Gate welcome screen is not a menu, and this rule does not reach it. Its bullets are read-only context introducing what this skill builds — they carry no numbers and selecting one is impossible. They precede the AskUserQuestion menu and never replace it; both must appear. Emit them as markdown bullets exactly as written. The rule above governs the choice, not the introduction to it.
• AskUserQuestion renders at most 4 options per question. Any beyond the fourth are silently dropped by the client — a hard runtime cap, not a display setting. Split into a second grouped question rather than adding a fifth option.
• Every AskUserQuestion question object must include multiSelect explicitly (true or false) — including every question in a batched multi-question call. Omitting it is a schema validation error (Invalid elicitation input: multiSelect ... invalid_type), not an optional default. This has caused repeated failed calls in production — always set it, even for single-question calls.
• Pass each question, label, and description as plain text — no markdown, no emoji, no line breaks. The chooser renders the string verbatim, so any markup shows as literal characters. The **…** in the tables below is doc formatting only; strip it when you pass the value.
• Currency — derive from the data, never default to USD. The reports enforce this themselves; do not pre-empt them with a guess.

Response style

The reader is an accountant working in a spreadsheet, and the deliverable is the workbook. What helps them is the content this skill defines — the welcome screen, the menu, the questions, the finished tabs. Running commentary on gate progress and tool plumbing doesn't help that reader, and it pushes the content they came for further down the screen. So this skill's replies are the content itself.

The Router Gate runs first, before any MCP tool call. If the prompt is generic/ambiguous, the welcome screen + AskUserQuestion menu is the very first thing you emit — the user sees what this skill can do while Gate 0 and Gate 0.5 haven't even started. If the prompt names a specific report, determine <REPORT> from the routing table and move straight to Gate 0.

Once <REPORT> is set, Gate 0 and Gate 0.5 run without commentary. Once both pass, load the matched report and let its own gates drive the rest — do not re-implement or pre-run its logic here.

The Carta connector asks for its welcome tool as your first action on connecting; that directive starts at Gate 0. It governs the order of Carta tool calls — welcome before set_context, list_contexts, call_tool, or fetch — and the welcome screen is text, not a tool call. Emit the screen, then call welcome at Gate 0. If the handshake already fired, the screen is still owed: a connection banner is not the welcome screen, and no tool emits the screen — you write it.

Never run a generic tool search — not tool_search_tool_bm25, not to discover a server prefix, not anywhere in this skill. Gate 0 explains why; the ban applies from the moment this skill loads.

The user-facing text this skill produces is:

1. The Router Gate welcome screen + AskUserQuestion menu — when the prompt is generic/ambiguous, before Gate 0 begins.
2. A STOP-row redirect — when the request is out of scope (during the Router Gate).
3. Firm disambiguation via AskUserQuestion — when multiple firms match (during Gate 0).
4. The entity picker and the period question via AskUserQuestion (during Gate 0).
5. The surface / workbook question and the new-file notice (during Gate 0.5).
6. The one-line build announcement before the report is loaded.

Everywhere else in the routing and gate sequence, the reply is the tool call. Commentary that adds nothing for this reader:

• Progress announcements — "Now I have the Carta MCP tools", "Proceeding with Gate 0", "Now checking…", "Now calling…".
• Server or prefix details — "Server prefix is <X>", "Found server <X>".
• Summaries of a tool result — "Context set to…".
• Tool inventories — "Only <X> surfaced", "this server lacks list_contexts". list_contexts and call_tool load lazily and are often absent from the visible tool list; that is expected and they still work, so there is nothing here worth reporting.
• MCP tool or command names in user-facing text, including inside an error message or a request for help. Describe what you need in plain English ("I can't reach your Carta connector") and name only the connector the user sees in their settings.
• A description of a step in place of the step. When a step calls for text, that text is the reply: a sentence about the welcome screen leaves the reader without the screen. Third-person narration ("The user asked for…", "no specific report was named"), narrated deliberation ("Let me check the args"), and announced intent ("I need to show the welcome menu") are all this same failure.
• Internal tool errors — context_snip failures, compression notes, and similar plumbing.

Entry mode — fresh session vs. chained skill

Check whether these context variables are already set from an earlier call in the same session:

• <SERVER> — connected Carta MCP server prefix
• <FIRM_NAME> and <FIRM_UUID> — the resolved firm
• <ENTITY_SCOPE> — the entities to include (all, or an explicit list)
• <PERIOD_START> and <PERIOD_END> — the reporting period
• <RUNTIME> — excel-addin or local-file
• <TARGET_FILE> — the workbook to write into (local-file runtime only)
• <REPORT> — previously routed report, if re-entering from a next-step menu. One of pnl, balance-sheet, trial-balance, or all (all three, built in one pass — see "Build" below).

Step order is always: Router Gate → Gate 0 → Gate 0.5 → Dispatch. The Router Gate carries its own user-facing output — the welcome screen and the menu — and that output lands before the connector's welcome call, not after.

If <REPORT> is already set: skip the Router Gate — the user already picked or named a report in this session.

If <SERVER>, <FIRM_NAME>, and <FIRM_UUID> are all set: skip the firm resolution in Gate 0 and reuse them. Do not ask "which firm?" when it is already established from the skill the user just ran.

If <RUNTIME> and <TARGET_FILE> are set: skip Gate 0.5.

Router Gate — Determine the right report

STOP rows — handle before routing: check these first. A match here means the request is out of scope for this skill entirely — redirect and stop; do not proceed to Gate 0.

Name the target skill verbatim in every redirect. Write the identifier out in full — carta-investors:carta-explore-data, carta-investors:carta-manco — rather than paraphrasing it into a description of what it does ("I'll route it through the data warehouse", "the budgeting tools"). A paraphrase leaves the user with nothing to act on; the identifier is the one piece of the message they can actually use. One sentence of plain-English context around it is fine.

If none of the rows above match, continue below.

Infer the report from the user's prompt. Do not ask the user to name a report by its technical name. Two paths only:

• Specific prompt — matches a row in the table below. Set <REPORT> immediately, do not dispatch yet, and do not emit the welcome screen or call AskUserQuestion. Proceed straight to Gate 0.
• Generic / ambiguous prompt — matches no row, or the skill was invoked with no specific task. Fall through to the Customer Intent Framework above before giving up and showing the welcome screen and AskUserQuestion menu.

Route rows — classify and proceed to Gate 0:

If ambiguous (prompt matches no row, or skill was invoked with no specific task), emit a welcome screen first, then ask via AskUserQuestion.

Welcome screen — output before calling AskUserQuestion, and before any Gate 0 tool call. It opens the reply; nothing precedes it — not a sentence about it, not the connector handshake. Emit the block below exactly as written. It is already complete as it stands: only the headline and the closing italic line vary, and both carry their default inline, so emitting it verbatim is always correct. Format rule: put each report on its OWN line as a markdown bullet (- ). Do NOT merge them into one paragraph — run together they render as an unreadable wall of text in Claude for Excel.

Ready to build a consolidating financial report.

Here's what I can build for you:

• Consolidating P&L — Income statement across your entities, with a detail tab and a one-page executive summary. Can also break spend down by department or reporting tag.
• Consolidating balance sheet — Assets, liabilities, and equity side by side by entity, with a firm-wide total.
• Consolidating trial balance — Every account balance, as a flat consolidated tab plus a by-entity tab.
• All three, back to back — Builds all three in one pass — I'll resolve the firm, entities, and period once and reuse them for every report.

I can write straight into an open Excel workbook, or build a standalone .xlsx file — whichever you're working with.

Headline variant: swap the first line for *"Connected to [FIRM] via Carta Fund Admin."* only when a real firm name is already known from a chained call this session. Never emit a bracketed placeholder.

The closing italic line tells the user where the workbook comes from before they pick a report — never drop it. Swap it for the matching variant when the surface is already obvious from what is in front of you (Excel add-in tools present, or a file path/attachment in the prompt — the same signals Gate 0.5 uses, read statically, with no tool call and no question here):

• excel-addin → I'll add new tabs to the workbook you have open — or update them if they're already there.
• local-file → Attach an existing workbook and I'll update it, or I'll build a new .xlsx file and give you the path when it's done.
• Unclear → keep the neutral line as written above. Gate 0.5 asks explicitly later.

Never ask the runtime question here — that belongs to Gate 0.5. This choice picks the wording of one line, nothing more.

Then ask (one AskUserQuestion, 4 options):

Which consolidating report can I build for you?

Store <REPORT> from the chosen option, then proceed to Gate 0. Do not dispatch yet — dispatch happens only after Gate 0.5.

Gate 0 — Carta MCP environment, firm, entity scope, and period

Scan the tools available in the conversation for any matching mcp__*__welcome. Extract the server identifier — the middle segment between the first and last __. Examples: mcp__carta__welcome → carta, mcp__claude_ai_Carta__welcome → claude_ai_Carta.

If none found: tell the user no Carta MCP is connected and stop. If exactly one found: call mcp__<SERVER>__welcome(_instrumentation_v2={"skills": ["carta-investors:carta-consolidating-financial-reports"]}) to verify. This is <SERVER>. If multiple found: ask which to use via AskUserQuestion. Default to carta (production) if present. Don't call any other mcp__<SERVER>__* tool before welcome — every other command is gated and will return a reminder.

The five Carta tools load lazily — a short tool list is not a missing tool

Derive <SERVER> from the server name as shown above. After that, the five suffixes welcome, set_context, list_contexts, call_tool, fetch are exhaustive for every Carta MCP server regardless of environment. Call mcp__<SERVER>__<suffix> directly.

list_contexts and call_tool load lazily. They will often NOT appear in your tool list, and that is expected — they still exist and still work. Only welcome, set_context, and fetch are reliably visible up front; a server may also expose extras like get_current_user or mutate that this skill does not use. Never treat a tool's absence from the visible list as evidence it is unavailable, and never tell the user a Carta tool "doesn't exist" or that the connector is missing a capability. Just call it.

Never search for these tools. Do not run tool_search_tool_bm25 under any circumstances — not to discover the prefix, not to find list_contexts or call_tool, not for anything. A BM25 search returns only the eagerly-loaded subset, so it will appear to prove the lazy tools are missing when they are not. That false negative is the trap: it leads to abandoning the run and telling the user to re-authenticate a working connector. If you genuinely must search, use the connector's own mcp__<SERVER>__search_tools, never the generic search.

If a call to one of the five genuinely errors, handle it from the error table at the end of this skill — do not infer from a tool listing.

Instrumentation — every Carta MCP call carries _instrumentation_v2

Pass it on every welcome, set_context, list_contexts, call_tool, and fetch call, exactly as written in the examples below:

_instrumentation_v2={"skills": ["carta-investors:carta-consolidating-financial-reports"]}

In Claude Code a hook injects this automatically, but in Claude for Excel there are no hooks — the call is rejected outright if the skill does not supply it. So always send it explicitly; sending it twice is harmless.

Add "model": "<the running model id>" only if you can actually introspect the running model. If you cannot, omit the key — never guess a model id, and never substitute a family name like claude-opus.

Resolve the firm

If the user named a firm → mcp__<SERVER>__list_contexts(firm_name="<entity>", _instrumentation_v2={"skills": ["carta-investors:carta-consolidating-financial-reports"]}) → disambiguate via AskUserQuestion if multiple → mcp__<SERVER>__set_context(firm_id=<FIRM_UUID>, _instrumentation_v2={"skills": ["carta-investors:carta-consolidating-financial-reports"]}).

Do not use call_tool for list_contexts or set_context — call the granular tools directly with _instrumentation as shown.

If the user named no firm, ask for it — but phrase it as the firm or management company, and say that the entity choice comes next, so the user isn't left thinking the firm answer decides their scope:

Which firm should I pull this from? I'll ask which funds and entities to include right after.

Resolve the entity scope

Always run this step. These are consolidating reports — the point is that they span more than one entity, so the user must be able to say which ones. Defaulting silently to "everything under the firm" is what makes the output surprising.

Load references/entity-picker.md and follow it. It lists the entities under the firm, groups them (management company, funds, SPVs), and builds the picker within the 4-option AskUserQuestion cap.

Store <ENTITY_SCOPE> as either all or the explicit list of entity names and UUIDs the user chose. Pass it to the report at Dispatch.

Never phrase this step as "which firm?" when what you mean is "which funds?" Customers read "firm" as the whole management company and are then surprised when a chosen fund's numbers don't appear on their own. Ask about funds and entities explicitly.

Resolve the reporting period

Collect the period before dispatching, so the report doesn't ask again.

• If the user gave an explicit period — a month, a quarter, a year, or a start and end date — use it.
• If they didn't, ask via AskUserQuestion, and offer a date range, not just a single month. An accountant asking for a P&L usually means a span of time; offering only one month is the most common source of "this isn't what I expected" on these reports.

What period should the report cover?

Store <PERIOD_START> and <PERIOD_END> as YYYY-MM-DD.

The P&L renders a period block covering <PERIOD_START>→<PERIOD_END> alongside a year-to-date block through <PERIOD_END>. The balance sheet and trial balance are as-of statements — they use <PERIOD_END> only, and <PERIOD_START> is ignored. Say so in one clause if the user picked a range for one of those two:

A balance sheet is a snapshot, so I'll build it as of [PERIOD_END].

DWH param-name traps (for the report to apply, not this skill): dwh:execute:query takes sql: not query:. dwh:get:table_schema takes table_name: not table:. format accepts "ndjson" / "markdown", not "csv".

Gate 0.5 — Surface and target workbook

Every report in this skill writes into an Excel workbook. Where that workbook comes from depends on the surface, and getting this wrong is the most common reason a run appears to do nothing at all.

If <RUNTIME> was already inferred before the welcome screen (Router Gate), reuse it here without re-detecting. That early pass uses the exact same signals below, so it never needs a second look — this step only runs the detection fresh when <RUNTIME> is still unset (a specific-prompt request skipped the welcome screen and never inferred it, or the earlier pass left it genuinely unclear).

Set <RUNTIME>:

• excel-addin — Claude for Excel. Signals: the user refers to "this workbook", "the open spreadsheet", or a tab, with no file path; or Excel add-in tools are available in the conversation.
• local-file — Claude Code, Cowork, or the desktop app. Signals: the user supplied a file path (~/Downloads/Financials.xlsx), attached a file, or asked to "create a new file"; or no Excel add-in tools are available.

If genuinely unclear, ask via AskUserQuestion: "Are you working in Excel through Claude for Excel, or with a local .xlsx file?"

If <RUNTIME> is excel-addin

The workbook in front of the user is the target. Set <TARGET_FILE> = null and proceed — the report's own destination gate handles the open-workbook cases (no workbook, empty workbook, workbook with existing tabs).

If <RUNTIME> is local-file

Resolve <TARGET_FILE>:

1. If the user attached a spreadsheet or named a path, set <TARGET_FILE> to it. The report adds its tabs to that file.

2. If not, set <TARGET_FILE> = null and tell the user, in one sentence, before any work starts:

   No spreadsheet attached, so I'll build this as a new .xlsx file and give you the path when it's done. If you'd rather I add these tabs to an existing workbook, attach it or give me the file path.

   Do not stop for an answer — proceed. The user can redirect if they want to.

Surface guidance — say this at most once, and only when it helps. These reports were built for Claude for Excel, where the tabs land directly in the workbook already open in front of the user. They work in Cowork, the desktop app, and Claude Code too; the only difference is where the workbook comes from. If the user seems unsure which surface to use:

These build fastest in Claude for Excel, where I write straight into your open workbook — but I can also produce the file here for you to open in Excel.

Do not repeat this, and do not lead with it when the user already named a report and a period.

Build — load the report and execute it inline

Both gates have passed.

<REPORT> is pnl, balance-sheet, or trial-balance

Emit exactly one line:

Building your consolidating [P&L | balance sheet | trial balance] for [FIRM_NAME].

Then load the matched report and follow it exactly:

Follow the loaded file verbatim from its own gates. Do not reconstruct its logic from memory — each one carries exact SQL, Excel number-format strings, column maps, and approval gates that must not be paraphrased.

<REPORT> is all

Build all three, one after another, into the same target workbook. Gate 0 and Gate 0.5 already ran exactly once for the whole batch — that single firm, entity-scope, period, and destination resolution is what "all three" saves over running each report on its own. Each report still runs its own DWH data pull; see "Why the data pull isn't shared" below.

Emit exactly one line before the batch starts:

Building your consolidating P&L, balance sheet, and trial balance for [FIRM_NAME] — one setup, three tabs.

Then, in this order — P&L, balance sheet, trial balance — emit that report's own one-line announcement (as in the single-report case above) and load it:

1. read_skill(file_path="references/pnl.md")
2. read_skill(file_path="references/balance-sheet.md")
3. read_skill(file_path="references/trial-balance.md")

Follow each file verbatim from its own gates, exactly as in the single-report case — including that report's own approval gate before it writes. Do not batch all three approvals into one prompt; let each report ask separately, back to back. If one report fails or the user declines its approval gate, say so in one sentence and continue to the next report rather than aborting the whole batch.

Why the data pull isn't shared. The three reports query different account-type ranges and different time semantics — the P&L sums two date windows (period and YTD) over accounts >= '4000', the balance sheet takes a single cumulative as-of balance over accounts '1000'–'3999', and the trial balance computes a beginning balance plus period debits/credits over every account. No single query shape serves all three without rewriting every report's data-pull gate, which is out of scope here. "One setup" refers to the firm/entity/period/destination resolution shared via Gate 0 and Gate 0.5, not to the DWH query itself.

The gates you already ran are already satisfied. Gate 0 and Gate 0.5 have set <SERVER>, <FIRM_NAME>, <FIRM_UUID>, <ENTITY_SCOPE>, <PERIOD_START>, <PERIOD_END>, <RUNTIME>, and <TARGET_FILE>. Every report reads these variables under the same names in its own "Entry mode" section and skips the gates whose inputs are set. Do not re-resolve the firm, re-ask the entity scope, or re-ask the period — the report begins at the first gate whose inputs are still missing, which is normally its data pull.

The report's own sub-references resolve inside this skill's tree (references/pnl/schema.md, references/balance-sheet/formatting.md, and so on) and its Excel branding asset resolves from this skill's assets/. Use those paths as written in the loaded file — do not rewrite them back to another skill's directory.

If Something Goes Wrong

Out-of-scope topics are handled proactively by the STOP rows in the Router Gate — they never reach this table. This table covers issues that surface mid-flow.

References

Report entry points

• references/pnl.md — consolidating P&L: detail tab + executive Summary tab, optional tag-view
• references/balance-sheet.md — consolidating balance sheet: Assets / Liabilities / Equity by entity
• references/trial-balance.md — consolidating trial balance: flat consolidated tab + by-entity tab

<REPORT> = all has no reference file of its own — it loads the three files above in sequence from the "Build" section, reusing the firm/entity/period/ destination resolved once in Gate 0 and Gate 0.5.

Shared by this skill

• references/entity-picker.md — list the entities under a firm, group them, and build the picker within the AskUserQuestion option cap
• references/local-file-output.md — producing a workbook outside Claude for Excel (runtime gate, destination, Office.js→operations translation, verification readback, path-based closing summary)

Maintainers only — never loaded at runtime

• docs/architecture-notes.md — why this skill is built the way it is: the inline-orchestrator pattern, the mirroring obligation against the standalone report skills, the self-contained-package rule, and the publish cutover. Read it before restructuring the skill or re-mirroring a report body. Do not load it while serving a user — it carries no runtime instruction, and docs/ is stripped from the published plugin.

P&L sub-references (loaded by references/pnl.md)

• references/pnl/schema.md — journal-entries column contract, sign conventions, period semantics, entity filtering
• references/pnl/section-map.md — keyword table assigning expense accounts to sections
• references/pnl/formatting.md — number formats, column widths, row styling
• references/pnl/branding-and-header.md — Carta branding, header layout, powered-by asset
• references/pnl/summary-tab.md — executive Summary tab layout and cross-sheet formula links
• references/pnl/budget-fetch.md — pulling stored budget figures for the Budget columns
• references/pnl/fill-budget-columns.md — writing budget figures into the period columns
• references/pnl/tag-view.md — Actuals broken down by reporting-tag category

Balance-sheet sub-references (loaded by references/balance-sheet.md)

• references/balance-sheet/schema.md — journal-entries column contract and sign conventions
• references/balance-sheet/formatting.md — number formats and layout
• references/balance-sheet/branding-and-header.md — Carta branding and header layout

references/trial-balance.md is self-contained and has no sub-references.

Assets

• assets/powered_by_carta.png and assets/powered_by_carta.b64.txt — this skill's own copy of the branding asset, shared by all three report bodies. The base64 sidecar is read package-relative at Excel runtime via blobs.getText("assets/powered_by_carta.b64.txt"), so it has to be a real file in this skill, not a path rewrite.