The plugin system

How it works

This is a largely Claude generted explanation of the tools Anthropic released for financial services. The financial-services release ships as a Claude Code plugin marketplace at github.com/anthropics/financial-services. Plugins inside it cluster around three vocabulary items.

Skills

A skill is a self-contained workflow document — a markdown file with a YAML (a human readable language) header. Its descriptionfield is the trigger: Claude scans it for keywords (“update thesis for X”, “find ideas”, “DCF model”) and auto-loads the skill into context when a conversation matches. The body is the workflow: numbered steps, output specifications, guardrails (“don’t take shortcuts”, “cite every number”). Skills don’t execute — they are prompt fragments that come to the fore when relevant.

Commands

Commands are slash-shortcuts that wrap a skill behind a name. /screenis three lines: a description, an argument hint, and “Load the idea-generation skill and run a quantitative screen.” The user types the command; the skill loads.

Named agents

An agent is a system prompt plus a list of skills it delegates to. The earnings-revieweragent says, in effect: “You are a senior research associate. Given a ticker, you produce three artifacts (model, note, variance table). Workflow: pull the print, invoke earnings-analysis, invoke model-update, invoke audit-xls, invoke morning-note. Guardrails: cite every number; never publish.” The agent is a thin orchestration layer; the skills do the work.

The discipline rules that recur

Reading across the skills, four conventions show up repeatedly:

• Formulas over hardcodes. Every projection cell in a financial model must be a live formula referencing input drivers, not a pre-computed number pasted in.
• Step-by-step verification. Show inputs → confirm → show projections → confirm → show valuation → confirm. Never end-to-end without checkpoints.
• Cite every number. If a figure cannot be sourced, mark it [UNSOURCED]. Cell comments with source on every input.
• No “helpful” extras.The initiating-coverage skill explicitly forbids creating “completion summaries, executive summaries, next-steps documents.” Deliver only what’s specified.

Loading order at session start

When Claude Code starts a session, plugin manifests are read; MCP servers (data partners like Daloopa, FactSet, S&P/Kensho) connect or fail silently if no auth is configured; skill descriptions are pre-indexed for trigger matching; commands appear under /as the user types. When a skill triggers, its full body is loaded into context for that turn. Agents are invoked when their description matches a request — they replace the conversation’s system prompt with their own and orchestrate from there.