audit-xls
Audit spreadsheet, formula error detection, hardcoded cell finder, cross-sheet reference audit, workbook auditor, Excel model audit, financial model QA, spreadsheet review, cell dependency trace, formula integrity check
Install
npx skills add https://github.com/agentii-ai/agentii-investment-intelligence/tree/main/plugins/vertical-plugins/models-and-pitches/skills/agentii/audit-xls
claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install agentii-ai-agentii-investment-intelligence@llmmart
git clone https://github.com/agentii-ai/agentii-investment-intelligence.git
The skills CLI installs just this skill, for any of its supported agents. Claude Code installs the whole agentii-ai/agentii-investment-intelligence collection as a plugin from our marketplace. Git is the plain clone.
Skill manifest
audit-xls
Triggers
- Audit spreadsheet
- formula error detection
- hardcoded cell finder
- cross-sheet reference audit
- workbook auditor
- Excel model audit
- financial model QA
- spreadsheet review
- cell dependency trace
- formula integrity check
Defaults
| Parameter | Default Value | Rationale |
|---|---|---|
| ticker | (required) | Stock symbol to analyze |
| lookback_quarters | 1 | Standard lookback for this skill type |
Methodology
1. Retrieval Scope
This skill operates with retrieval_scope: simple_lookup. It uses only profile/entity metadata tools — no document or XBRL retrieval at scale.
2. Retrieval Strategy
Follows the retrieval strategy decision tree in contracts/retrieval.md. Primary branch: (d) Simple Lookup. Resolve the canonical ticker first (exact → fuzzy alias → share-class) before any data call.
3. Temporal Scope
Default lookback: 1 fiscal quarter(s); maximum: 1. The default balances recency against the trend window this analysis requires.
4. Tool Allowlist
Per frontmatter allowed_tools:
search_companies— ticker resolution + company context (entity-alias fuzzy match)get_company_financials— consolidated IS/BS/CF highlightsget_calculation_tree— XBRL calculation linkbase (weights)validate_calculation— XBRL calc-consistency validationlist_sources— used by this skill per the retrieval strategy
5. Protocol
- Pre-flight:
get_company_fiscal_calendar/{ticker}thenget_ticker_coverage/{ticker}. - Lookup:
get_company_profile/{ticker}/get_entity_knowledgefor the requested metadata field(s). - Output: write the deliverable per
## Output File, then append toagentii.md.
Deliverable Chain
Inputs → Build → Validate → Output → Next
- Inputs: resolved ticker + structured facts (
search_xbrl_facts,get_company_financials) and any filing pages from the three-layer protocol. - Build: perform the formula / hardcoded-cell / cross-sheet-reference audit on the input workbook and write the
.mdaudit report per## Output Structure. - Validate: run the
## Validation Gatesbelow. - Output: write the artifact path per
## Output File. - Next: append to
agentii.md; hand off to a downstream pitch/review skill if requested.
Validation Gates
- **calculation arc cross-validation **: workbook computed totals verified against
gold.xbrl_calculationsweights. Compareget_calculation_tree(accession_number)expected values against workbook formulas. Flag discrepancies ≥1% of parent concept value. If failed: If material discrepancy (≥5%): refuse delivery. If minor (1-5%): flag in audit findings withseverity: warning. - hardcoded cell detection: zero hardcoded values in cells tagged as formulas. Use
xlsx_audithardcoded-count output. If failed: If hardcoded_count > 0: refuse delivery with audit report listing each hardcoded cell location. - cross-sheet reference integrity: all cross-sheet references resolve to valid cell ranges. If failed: If broken references found: refuse delivery with broken reference map.
- tool diversity: distinct MCP tools used in this invocation >=
min_tool_diversity(3). If failed: flag as depth-insufficient in Coverage Gaps.
Output File
Write the final deliverable to {ticker}/{YYYY-MM-DD_HHMM}_audit-xls_{affix}.md .
Output Structure
- Executive Summary (≤200 words) — headline conclusions for the analysis.
- Data Sources — filings + structured endpoints used, with
{ticker} {citation_id} page<N>citations. - Analysis — the core findings, tables, and commentary for this dimension.
- Key Metrics — the quantitative results with QoQ/YoY context where relevant.
- Coverage Gaps & Citations — data not retrievable + citation index.
Citations & memory: follow contracts/citation-and-memory.md — ≥1 citation per 200 words; every material fact, table row, and metric is immediately followed by its inline clickable https://agentii.ai/v/{ticker}/{citation_id}/{N} link; a bottom Citations section provides a non-duplicative roll-up index; the closing TUI reply includes a compact Key Citations list (headline 5–10 facts) of clickable /v/ URLs; and append the run to agentii.md per contracts/agentii-md-schema.md.
Preflight
Run the canonical pre-flight sequence — MCP health probe, ticker resolution, workspace style.md override, memory load, and coverage check. See contracts/preflight.md.
Include the X-Agentii-Trace header on every tool call per contracts/x-agentii-trace-header.md.
Memory & Snapshot
- Memory load (pre-flight): load prior workspace context for the ticker before retrieval — see
contracts/memory-load.md. - Structured output frontmatter: emit the FR-090 block (
key_metrics,conclusions,facts_count,deducted_count,views_count,citation_count) percontracts/output-frontmatter-schema.md. - Snapshot synthesis: after writing the deliverable, update the two-tier snapshot and classify findings as
[FACT]/[DEDUCTED]/[VIEW]— seecontracts/snapshot-synthesis.md. - Session archival: record the run under
sessions/{YYYY-MM-DD}/and updatesessions/INDEX.mdpercontracts/session-format.md.
Final Summary (TUI)
End the closing chat reply with a compact Key Citations list (headline 5–10 facts), each a clickable https://agentii.ai/v/{ticker}/{citation_id}/{N} link, so the user can cmd+click straight to the exact SEC page. See contracts/citation-and-memory.md.
Error Handling
| Error | Action |
|---|---|
| Ticker not found | Suggest checking spelling or trying list_coverage |
| No data available | Flag in Coverage Gaps, proceed with available data |
| API key invalid | Direct user to agentii.ai/api-keys |
| MCP server unreachable | Retry once; if persistent, halt with AGENTII_MCP_UNREACHABLE |
Files (agentii-investment-intelligence)
-
references
-
formula-sheet.md 128 B
# Audit Xls — Canonical Formulas See `## Deliverable Chain` and `## Validation Gates` in SKILL.md for the complete workflow. -
institutional-defaults.md 106 B
# Audit Xls — Institutional Defaults See `## Defaults` table in SKILL.md for default parameter values. -
modes.md 709 B
# audit-xls — Analyst Mode Definitions Derived from the skill's own methodology structure (scripts/mode_backfill.py, spec 046 M1). ### Mode: triggers **Objective**: Triggers analysis per the skill's methodology (see SKILL.md sections). ### Mode: defaults **Objective**: Defaults analysis per the skill's methodology (see SKILL.md sections). ### Mode: methodology **Objective**: Methodology analysis per the skill's methodology (see SKILL.md sections). ### Mode: retrieval-scope **Objective**: Retrieval Scope analysis per the skill's methodology (see SKILL.md sections). ### Mode: retrieval-strategy **Objective**: Retrieval Strategy analysis per the skill's methodology (see SKILL.md sections). -
validation-checklist.md 475 B
# Audit Xls — Validation Checklist Expanded version of `## Validation Gates` with per-gate remediation steps. 1. **hardcoded_count**: reported for every tagged category - *Remediation*: If any category unreported: refuse delivery. 2. **formula_trace**: lists all cross-sheet references - *Remediation*: If trace empty but formulas exist: flag as gap. 3. **standard agentii citations**: checked on all cell comments - *Remediation*: If non-conforming: list violations.
-
-
SKILL.md 6.5 KB
--- name: audit-xls multi_ticker_semantics: single_target description: Audit spreadsheet, formula error detection, hardcoded cell finder, cross-sheet reference audit, workbook auditor, Excel model audit, financial model QA, spreadsheet review, cell dependency trace, formula integrity check temporal_scope: default_quarters: 1 max_quarters: 1 description: "Typical lookback: 1 quarters, max: 1" allowed_tools: - search_companies - get_company_financials - get_calculation_tree - validate_calculation - list_sources - xlsx-read retrieval_scope: simple_lookup min_tool_diversity: 3 --- # audit-xls ## Triggers - Audit spreadsheet - formula error detection - hardcoded cell finder - cross-sheet reference audit - workbook auditor - Excel model audit - financial model QA - spreadsheet review - cell dependency trace - formula integrity check ## Defaults | Parameter | Default Value | Rationale | |-----------|---------------|-----------| | ticker | (required) | Stock symbol to analyze | | lookback_quarters | 1 | Standard lookback for this skill type | ## Methodology ### 1. Retrieval Scope This skill operates with `retrieval_scope: simple_lookup`. It uses only profile/entity metadata tools — no document or XBRL retrieval at scale. ### 2. Retrieval Strategy Follows the retrieval strategy decision tree in `contracts/retrieval.md`. Primary branch: **(d) Simple Lookup**. Resolve the canonical ticker first (exact → fuzzy alias → share-class) before any data call. ### 3. Temporal Scope Default lookback: 1 fiscal quarter(s); maximum: 1. The default balances recency against the trend window this analysis requires. ### 4. Tool Allowlist Per frontmatter `allowed_tools`: - `search_companies` — ticker resolution + company context (entity-alias fuzzy match) - `get_company_financials` — consolidated IS/BS/CF highlights - `get_calculation_tree` — XBRL calculation linkbase (weights) - `validate_calculation` — XBRL calc-consistency validation - `list_sources` — used by this skill per the retrieval strategy ### 5. Protocol 1. **Pre-flight**: `get_company_fiscal_calendar/{ticker}` then `get_ticker_coverage/{ticker}`. 2. **Lookup**: `get_company_profile/{ticker}` / `get_entity_knowledge` for the requested metadata field(s). 3. **Output**: write the deliverable per `## Output File`, then append to `agentii.md`. ## Deliverable Chain **Inputs** → **Build** → **Validate** → **Output** → **Next** 1. **Inputs**: resolved ticker + structured facts (`search_xbrl_facts`, `get_company_financials`) and any filing pages from the three-layer protocol. 2. **Build**: perform the formula / hardcoded-cell / cross-sheet-reference audit on the input workbook and write the `.md` audit report per `## Output Structure`. 3. **Validate**: run the `## Validation Gates` below. 4. **Output**: write the artifact path per `## Output File`. 5. **Next**: append to `agentii.md`; hand off to a downstream pitch/review skill if requested. ## Validation Gates 1. **calculation arc cross-validation **: workbook computed totals verified against `gold.xbrl_calculations` weights. Compare `get_calculation_tree(accession_number)` expected values against workbook formulas. Flag discrepancies ≥1% of parent concept value. *If failed*: If material discrepancy (≥5%): refuse delivery. If minor (1-5%): flag in audit findings with `severity: warning`. 2. **hardcoded cell detection**: zero hardcoded values in cells tagged as formulas. Use `xlsx_audit` hardcoded-count output. *If failed*: If hardcoded_count > 0: refuse delivery with audit report listing each hardcoded cell location. 3. **cross-sheet reference integrity**: all cross-sheet references resolve to valid cell ranges. *If failed*: If broken references found: refuse delivery with broken reference map. 4. **tool diversity**: distinct MCP tools used in this invocation >= `min_tool_diversity` (3). *If failed*: flag as depth-insufficient in Coverage Gaps. ## Output File Write the final deliverable to `{ticker}/{YYYY-MM-DD_HHMM}_audit-xls_{affix}.md` . ## Output Structure 1. **Executive Summary** (≤200 words) — headline conclusions for the analysis. 2. **Data Sources** — filings + structured endpoints used, with `{ticker} {citation_id} page<N>` citations. 3. **Analysis** — the core findings, tables, and commentary for this dimension. 4. **Key Metrics** — the quantitative results with QoQ/YoY context where relevant. 5. **Coverage Gaps & Citations** — data not retrievable + citation index. **Citations & memory**: follow `contracts/citation-and-memory.md` — ≥1 citation per 200 words; every material fact, table row, and metric is immediately followed by its inline clickable `https://agentii.ai/v/{ticker}/{citation_id}/{N}` link; a bottom **Citations** section provides a non-duplicative roll-up index; the closing TUI reply includes a compact **Key Citations** list (headline 5–10 facts) of clickable `/v/` URLs; and append the run to `agentii.md` per `contracts/agentii-md-schema.md`. ## Preflight Run the canonical pre-flight sequence — MCP health probe, ticker resolution, workspace `style.md` override, memory load, and coverage check. See `contracts/preflight.md`. Include the `X-Agentii-Trace` header on every tool call per `contracts/x-agentii-trace-header.md`. ## Memory & Snapshot - **Memory load** (pre-flight): load prior workspace context for the ticker before retrieval — see `contracts/memory-load.md`. - **Structured output frontmatter**: emit the FR-090 block (`key_metrics`, `conclusions`, `facts_count`, `deducted_count`, `views_count`, `citation_count`) per `contracts/output-frontmatter-schema.md`. - **Snapshot synthesis**: after writing the deliverable, update the two-tier snapshot and classify findings as `[FACT]`/`[DEDUCTED]`/`[VIEW]` — see `contracts/snapshot-synthesis.md`. - **Session archival**: record the run under `sessions/{YYYY-MM-DD}/` and update `sessions/INDEX.md` per `contracts/session-format.md`. ## Final Summary (TUI) End the closing chat reply with a compact **Key Citations** list (headline 5–10 facts), each a clickable `https://agentii.ai/v/{ticker}/{citation_id}/{N}` link, so the user can cmd+click straight to the exact SEC page. See `contracts/citation-and-memory.md`. ## Error Handling | Error | Action | |-------|--------| | Ticker not found | Suggest checking spelling or trying list_coverage | | No data available | Flag in Coverage Gaps, proceed with available data | | API key invalid | Direct user to agentii.ai/api-keys | | MCP server unreachable | Retry once; if persistent, halt with AGENTII_MCP_UNREACHABLE |
Comments (0)
Sign in to join the conversation.
Reviews (0)
No reviews yet.
No comments yet.