Case Study Orchestrator — Marketplace

Single-plugin Claude Code / Cowork marketplace hosting case-study-orchestrator.

What it is

A thin orchestrator that solves financial-services case studies end to end by routing through Anthropic's published financial-analysis, investment-banking, and private-equity plugins. The user invokes /solve-case-study from a case folder; the plugin inventories the materials, classifies the case shape (model-only, deck-only, or model+deck), and dispatches the smallest set of three scoped worker subagents required.

All methodology comes from skills preloaded into the workers via the skills: frontmatter field — full skill content is injected at subagent startup, not just the description. The plugin adds only what Anthropic's pre-built model-builder and pitch-agent subagents don't ship: Bash access (so the bundled python in xlsx-author / pptx-author can run) and case-study-specific I/O (reading prompts from PDFs or in-Excel instructions, handling yellow/blue template conventions, etc.). Anthropic's *-agent plugins are wired for live-Excel mode via mcp__office__excel_* and would not function in headless case-study runs.

Prerequisites

These plugins must be enabled for the worker preloads to resolve:

claude plugin marketplace add anthropics/claude-for-financial-services
claude plugin install financial-analysis@claude-for-financial-services
claude plugin install investment-banking@claude-for-financial-services
claude plugin install private-equity@claude-for-financial-services

Required versions: financial-analysis v0.1.1+, investment-banking v0.2.1+, private-equity v0.1.2+. Missing skills are skipped silently at subagent startup (a warning is written to the debug log).

Install

Claude Code (CLI)

claude plugin marketplace add ZaBrisket/case-study-orchestrator
claude plugin install case-study-orchestrator@case-study-marketplace

Cowork (Desktop)

Push this repository to GitHub.
In the Cowork desktop app, open the Customize menu and add this repo as a marketplace.
Install case-study-orchestrator from the marketplace.

Local development

claude plugin marketplace add /path/to/this/repo
claude plugin install case-study-orchestrator@case-study-marketplace

One-shot test without installing:

claude --plugin-dir ./plugins/case-study-orchestrator -p "Reply DONE."

Usage

cd path/to/case-folder
/solve-case-study

Or with an explicit folder argument:

/solve-case-study /absolute/path/to/case-folder

The orchestrator inventories the folder, classifies the case, and shows you a one-line dispatch plan before any subagent fires. A review gate fires before the deck-builder only when the deck depends on prior worker output.

Architecture

case-study-orchestrator/
├── .claude-plugin/marketplace.json
├── README.md                                       ← you are here
├── plugins/
│   └── case-study-orchestrator/
│       ├── .claude-plugin/plugin.json
│       ├── README.md                               ← architecture deep-dive
│       ├── skills/
│       │   └── solve-case-study/SKILL.md           ← orchestrator, /solve-case-study
│       ├── agents/
│       │   ├── case-digester.md                    ← read-only synthesizer
│       │   ├── case-modeler.md                     ← writes .xlsx (8 preloaded skills)
│       │   └── case-deck-builder.md                ← writes .pptx (8 preloaded skills)
│       └── scripts/                                ← bundled helpers (v0.2.0+)
│           ├── audit_lbo_template.py
│           ├── sensitivity_datatable.py
│           └── preserve_protected_sheets.py
└── tests/                                          ← regression harness (see tests/README.md)

Case shape	Workers dispatched
Model-only	`case-modeler`
Deck-only with narrative sources	`case-digester` → `case-deck-builder`
Model + deck (deck depends on model)	`case-modeler` → review gate → `case-deck-builder`
Independent sub-exercises (e.g., Republic Services pack)	`case-digester` ∥ `case-modeler` → `case-deck-builder`

Workers run in parallel when their inputs are independent. A single review gate fires before the deck-builder when the deck depends on prior worker output.

Output

Deliverables write to <case-folder>/_output/. Original input files are never modified. Each run suffixes .v2, .v3, etc. rather than overwriting existing deliverables.

Tests

A regression harness lives in tests/ — it runs the orchestrator against three known fixtures (Vortex / Republic Services / Edgewater) and validates the produced deliverables against frozen structural and numerical contracts (BS balance, sources-equal-uses tie, IRR/MoM within tolerance, slide count, slide titles, no error cells, etc.).

python tests/run_tests.py

Fixture materials are gitignored — the harness is structure-only until materials are added locally. See tests/README.md for the contract schema.

What's new in v0.2.0

Lessons from a real-world Summit LBO case run (May 2026):

Native Excel two-input data tables for sensitivity grids — the modeler now writes DataTableFormula via the bundled scripts/sensitivity_datatable.py.
scripts/audit_lbo_template.py extends financial-analysis:audit-xls with LBO templating-convention checks (hold-flag gating, S=U numeric tie, terminal debt, sensitivity-table format, in-workbook assumption labels).
scripts/preserve_protected_sheets.py keeps reference sheets (Answer Key, Variance) byte-identical to the source.
Label-vs-spec conflict surface — modeler flags conflicts rather than picking silently.
No autonomous v2 rewrites — the modeler's first deliverable is canonical; revisions happen on explicit user direction.
Versioning clarified — subsequent revisions: .FILLED.v2.xlsx (dot-separated, not _v2).
Working artifacts (recompute scripts, audit transcripts) go to <case-folder>/_output/_artifacts/, not alongside the deliverable.

See plugins/case-study-orchestrator/README.md for the architecture deep-dive and plugins/case-study-orchestrator/scripts/README.md for the helper APIs.

What's deliberately not included

No bundled MCP, no userConfig, no hooks. Self-contained and headless.
No external research in the default flow. Case studies typically forbid it.
No cross-case memory. Each dispatch is a fresh classification.
No dispatch to Anthropic's model-builder / pitch-agent subagents. Verified to be wired for live-Excel mode with no Bash access; they cannot run the bundled python in headless case-study runs.

License

MIT

case-study-orchestrator

Install to Claude Code

README