How panaversity/agentfactory-business-plugins works
1. This repository is a multi-plugin marketplace for Claude/AgentFactory-style enterprise agents. The top level acts as the distribution layer, while each subdirectory is a self-contained domain plugin with its own manifest, skills, evaluation suite, and sometimes commands, hooks, agents, and scripts.
2. The main architectural boundary is per-domain isolation. Each plugin package lives in its own folder, such as banking/, legal-ops/, islamic-finance/, idfa-financial-architect/, and other domain packs like sales-revops-marketing/, product-strategy/, operations-intelligence/, supply-chain/, hr-operations/, and innovation/. There is very little shared runtime code between them; the common pattern is structural, not a shared library.
3. Inside each plugin, the core execution unit is the skill. Skills are stored as skills/*/SKILL.md and usually paired with domain references and trigger files. These skill documents encode the behavior the agent should follow for a specific workflow, such as routing, drafting, reconciliation, compliance checks, or strategy generation.
4. Some plugins add commands as explicit user-facing entry points. These appear as Markdown command specs under commands/ and are used to expose named workflows like banking AML, legal intake, Islamic finance journaling, or sales sequence generation. Commands are the operational boundary between a generic chat input and a predefined domain workflow.
5. Several plugins also use agents as specialized roles. These are Markdown role definitions such as chief-of-staff, meeting intelligence, compliance monitor, contract intake, or spend intelligence. In architecture terms, these are the orchestration layer that decomposes a request into narrower sub-tasks handled by skills and reference content.
6. The references folders are the domain knowledge base. They hold static regulatory maps, jurisdiction overlays, legal rules, accounting standards, and other lookup tables. These are not executable components, but they materially shape routing and answer quality by constraining outputs to jurisdiction- or standard-specific rules.
7. The hooks layer provides validation and session-time enforcement where present. This is the guardrail layer: it can intercept plugin usage, validate routing, and prevent unsafe or out-of-domain behavior. In regulated plugins like banking, legal ops, and Islamic finance, hooks are part of the safety boundary rather than a convenience feature.
8. The repo has a strong evaluation and benchmarking subsystem. Each plugin carries evals/ or evals-workspace/, plus cases, goldens, graders, and generated outputs. The common pattern is: input case → skill/agent execution → output artifact → grading JSON. Many workspaces contain both “with_skill” and “without_skill” runs, which makes the evaluation boundary explicit.
9. Evaluation artifacts are organized as iterative workspaces. Folders like iteration-1, iteration-2, and iteration-3 hold benchmark snapshots, prompts, transcripts, outputs, timing, and grading results. This shows the repo is designed not just for shipping plugin content, but for continuously measuring and improving prompt/skill behavior over time.
10. The IDFA financial architect plugin is the most code-heavy subsystem. Unlike the mostly Markdown-based domain packs, it includes Python scripts, unit tests, Excel fixtures, generated models, and two plugin areas: financial-architect and idfa-ops. Its architecture centers on spreadsheet/model manipulation, auditing, and compliance checks, with Python used for validation and recalculation bridging.
11. The skill-creator tooling under .claude/skills/skill-creator is a meta-layer for producing and packaging skills. It includes scripts for validation, benchmark aggregation, report generation, and packaging, plus evaluation-viewer assets. This is repository-internal infrastructure for maintaining the plugin ecosystem itself.
12. CI/CD is handled through GitHub Actions workflows at the root, with separate eval and release pipelines for major plugins. This indicates the repo’s deployment unit is a plugin release, not an application build. The workflows likely validate skills/evals and then package or publish plugin artifacts.
13. The file layout suggests a clear source vs. generated-artifact split. Source lives in skills/, commands/, agents/, hooks/, references/, and scripts/; generated results live in workspaces, evals/results, description-optimization, and release/review artifacts. That separation is important for an architecture diagram because it distinguishes authored domain logic from benchmark outputs.
14. Overall data flow is: user intent enters via Claude plugin install/commands → plugin manifest activates the relevant skills/agents/hooks → reference content constrains the response → the plugin produces a domain-specific artifact or answer → eval harnesses grade outputs and feed iteration loops → CI/release packaging publishes updated plugin versions. The system is therefore a curated marketplace of modular, domain-specific AI behavior with strong evaluation and guardrail boundaries.
Main entry points: marketplace.json Marketplace · workflows CI workflows · .claude-plugin/plugin.json Banking · .claude-plugin/plugin.json Sales ops
Components 24
Distribution layer
Marketplace, repo-wide guidance, and release orchestration
- Marketplace · Top-level plugin distribution manifest
.claude-plugin/marketplace.json - CI workflows · Release and eval pipelines for major plugins
.github/workflows - Repo guide · Repository-wide operating guidance
CLAUDE.md
Meta tooling
Skill creation, packaging, and review infrastructure
- Skill creator · Repository tooling for authoring and packaging skills
.claude/skills/skill-creator/SKILL.md - Creator scripts · Validation, reporting, evaluation, and packaging scripts
.claude/skills/skill-creator/scripts - Eval viewer · HTML/Python review viewer for evaluation artifacts
.claude/skills/skill-creator/eval-viewer
Domain plugins
Self-contained Claude plugin packs by business domain
- Agentic office · Cross-functional workplace agent plugin
agentic-office/.claude-plugin/plugin.json - Banking · Regulated banking workflows and controls
banking/.claude-plugin/plugin.json - Legal ops · Legal operations and compliance workflows
legal-ops/.claude-plugin/plugin.json - Sales ops · Sales, revops, and marketing workflows
sales-revops-marketing/.claude-plugin/plugin.json - Product strategy · Product discovery, PRDs, prioritization, and retros
product-strategy/.claude-plugin/plugin.json - Ops intelligence · Operational monitoring, contracts, incidents, and metrics
operations-intelligence/.claude-plugin/plugin.json - HR ops · Hiring, knowledge, offboarding, and references
hr-operations/.claude-plugin/plugin.json - Supply chain · Procurement, vendors, logistics, and spend workflows
supply-chain/.claude-plugin/plugin.json - Islamic finance · Shariah-aware finance, screening, and journaling
islamic-finance/.claude-plugin/plugin.json - IDFA architect · Spreadsheet and model-driven financial architecture plugin
idfa-financial-architect/.claude-plugin/plugin.json - Bank router · Jurisdiction-aware banking routing and regulatory mapping
banking/skills/banking-global-router/SKILL.md - Sales router · Jurisdiction-aware sales and outreach routing
sales-revops-marketing/skills/sales-marketing-global-router/SKILL.md - Legal router · Jurisdiction-aware legal routing and constraints
legal-ops/skills/legal-global-router/SKILL.md - Islamic router · Jurisdiction and standards routing for Islamic finance
islamic-finance/skills/islamic-finance-router/SKILL.md - Financial architect · Core spreadsheet and model transformation skill
idfa-financial-architect/skills/financial-architect/SKILL.md - IDFA ops · Workbook audit, repair, and recalc bridge skill
idfa-financial-architect/skills/idfa-ops/SKILL.md
Evaluation system
Benchmarks, workspaces, and validation artifacts
- Eval harness · Plugin evaluation entrypoints and benchmark execution
- Workspaces · Iteration snapshots, outputs, grades, and timing artifacts
Connections 36
- Marketplace → Agentic office: distributes · Marketplace activates plugin packs
- Marketplace → Banking: distributes · Marketplace activates plugin packs
- Marketplace → Legal ops: distributes · Marketplace activates plugin packs
- Marketplace → Sales ops: distributes · Marketplace activates plugin packs
- Marketplace → Product strategy: distributes · Marketplace activates plugin packs
- Marketplace → Ops intelligence: distributes · Marketplace activates plugin packs
- Marketplace → HR ops: distributes · Marketplace activates plugin packs
- Marketplace → Supply chain: distributes · Marketplace activates plugin packs
- Marketplace → Islamic finance: distributes · Marketplace activates plugin packs
- Marketplace → IDFA architect: distributes · Marketplace activates plugin packs
- CI workflows → Eval harness: runs · CI runs plugin eval and release jobs
- CI workflows → Banking: releases · Banking release and eval pipeline
- CI workflows → Legal ops: releases · Legal release and eval pipeline
- CI workflows → Sales ops: releases · Sales release and eval pipeline
- CI workflows → Islamic finance: releases · Islamic finance release and eval pipeline
- CI workflows → IDFA architect: releases · IDFA release and eval pipeline
- Skill creator → Creator scripts: uses · Meta-skill delegates packaging and validation tasks
- Skill creator → Eval viewer: uses · Meta-skill generates and reviews eval reports
- Banking → Bank router: routes with · Regulatory router guides banking workflows
- Sales ops → Sales router: routes with · Outreach and compliance router guides workflows
- Legal ops → Legal router: routes with · Jurisdiction router constrains legal work
- Islamic finance → Islamic router: routes with · Standards and jurisdiction router constrains outputs
- IDFA architect → Financial architect: core skill · Primary spreadsheet and model transformation workflow
- IDFA architect → IDFA ops: ops skill · Audit and workbook repair workflow
- Banking → Eval harness: measures · Banking benchmarks validate routing and compliance
- Sales ops → Eval harness: measures · Sales benchmarks validate major workflows
- Product strategy → Eval harness: measures · Product strategy benchmarks validate authored artifacts
- Ops intelligence → Eval harness: measures · Ops intelligence benchmarks validate monitoring workflows
- HR ops → Eval harness: measures · HR benchmarks validate knowledge and document workflows
- Supply chain → Eval harness: measures · Supply chain benchmarks validate procurement workflows
- IDFA architect → Workspaces: iterates · Workspace iterations store outputs, grades, and timing
- Sales ops → Workspaces: iterates · Workspace iterations compare with-skill and without-skill runs
- Product strategy → Workspaces: iterates · Workspace iterations compare with-skill and without-skill runs
- Banking → Workspaces: iterates · Workspace iterations compare with-skill and without-skill runs
- HR ops → Workspaces: iterates · Workspace iterations compare with-skill and without-skill runs
- Repo guide → Marketplace: frames · Repo guidance frames plugin distribution and usage