Investment Banking Deck Checker
Find cross-slide number errors, unsupported claims, language issues, and presentation defects before client delivery.
The skill explicitly says it is read-and-report only and makes no deck edits, limiting external effects. It does not explain handling of sensitive client data, temporary-file retention, data flows, or user confirmation, so points are deducted for incomplete privacy, least-privilege transparency, and recovery safeguards.
It provides a structured workflow and a basic number-extraction script with file-not-found handling. However, there are no key-path tests; category grouping and a 5% tolerance can create false positives, while visual QC has no reproducible tool procedure or diagnostic failure feedback. Static evidence therefore warrants a score below 10.
The audience, use cases, and four review dimensions are clear. Boundaries for unsupported deck types, invalid inputs, semantic trigger exclusions, Chinese-language material, and practical visual-QC capability in chat are unspecified, so points are deducted for limited environment and trigger precision.
The documentation is reasonably layered, with terminology guidance, a report template, severity rules, an Apache-2.0 license, and repository-level CI maintenance signals. It lacks skill versioning, changelog, explicit ownership, dependency/install notes, FAQs, and troubleshooting; the script path also depends on the working directory.
The intended output covers number consistency, data-narrative alignment, language, and formatting, with slide attribution and severity prioritization, giving it clear practical value. Visual verification is underspecified, the script only supports extracted text, and the report is not clearly machine-ready. Under static calibration, the score is capped at 7 and reduced for incomplete direct usability evidence.
The source instructions, script logic, and report format are inspectable, and repository CI validates plugins and scans secrets. There is no committed test suite covering this skill's key paths, real execution evidence, or independent corroboration, so the score is limited to 4 and remains below the static maximum of 5.
- Do not treat the report as the sole pre-delivery control; qualified professionals should review figures, sources, narrative alignment, and visual issues.
- Confirm that the deck may safely be uploaded and establish rules for sensitive data, temporary files, and access permissions.
- Verify script paths, dependencies, and visual-QC tooling in the deployment environment; number categorization may produce false positives or negatives.
- Repository-level plugin-validation CI does not demonstrate that this skill's key paths are tested.
What does this skill do, and when should you use it?
This skill reviews pitch decks and client-ready presentations across four quality dimensions: number consistency, data-narrative alignment, IB language polish, and visual formatting. It reads every slide and preserves slide-level attribution for findings. A number-checking script normalizes monetary units and metric categories to identify conflicting values across slides. The workflow is read-only and produces a severity-ranked report rather than editing the deck.
Reads all slide text from an uploaded PPTX or a live PowerPoint add-in deck and saves it with slide markers; runs scripts/extract_numbers.py --check to normalize formats such as $M, $MM, and full dollar amounts and flag metric conflicts; verifies totals, percentages, growth rates, units, and reporting periods; tests whether trend and market-position claims are supported by the underlying data; scans for informal phrasing, contractions, exclamation points, vague quantifiers, and terminology drift; checks source citations, axis labels, typography, number and date formats, footnotes, disclaimers, overlaps, overflow, and contrast; reports findings as Critical, Important, or Minor.
- An investment banking analyst checks revenue, EBITDA, multiples, and margin figures across a pitch book before sending it to a client.
- A deal team verifies that revised charts still support the narrative written on earlier slides.
- A VP or execution team performs a final pass for missing sources, axis labels, footnotes, and disclaimers before delivery.
- A banking team runs a language review to remove casual phrasing, contractions, inconsistent terminology, and unsupported vague claims.
What are this skill's strengths and limitations?
- Covers four practical review areas: numbers, narrative, language, and visual formatting.
- Maintains slide-level attribution so reviewers can locate conflicting figures.
- Uses explicit severity levels and leads with delivery-blocking issues.
- Supports both PowerPoint add-in and chat workflows.
- The source provides no test results, accuracy metrics, or sample reports.
- It depends on scripts/extract_numbers.py and referenced files ib-terminology.md and report-format.md, whose contents are not provided.
- It reports issues but does not repair the presentation.
- Visual checks require access to the actual slide appearance; text extraction alone cannot fully detect overlap, overflow, or contrast problems.
- It requires Python, filesystem access, and a readable PPTX or live-deck environment.
How do you install this skill?
The skill is located at plugins/agent-plugins/pitch-agent/skills/ib-check-deck/ in anthropics/financial-services. In Cowork, open Settings → Plugins → Add plugin, add https://github.com/anthropics/financial-services, and select Pitch Agent; alternatively, upload a zip containing plugins/agent-plugins/pitch-agent/. For Claude Code, the documented commands are claude plugin marketplace add anthropics/financial-services followed by claude plugin install pitch-agent@claude-for-financial-services. The source does not document a standalone installation command for this individual skill.
How do you use this skill?
In a supported session, upload a PPTX or open a deck in the PowerPoint add-in, then ask: “Review this deck for number consistency and client-readiness,” or “Reconcile the figures across slides and check the formatting.” The skill reads the full deck and returns a read-only report. Its documented internal command is python scripts/extract_numbers.py /tmp/deck_content.md --check; the source does not describe a separate end-user command or interface.