capsa-cookbook

Capsa Agent Cookbook

A growing collection of recipes and installable skills that show how agents can use the Capsa MCP connector alongside other connected apps — email, calendars, docs — to complete real work.

What this cookbook is

The Capsa MCP connector is a capability API. This cookbook is the layer on top: practical, procedural recipes — packaged as installable skills — that teach an agent how to combine Capsa's context with the apps your team already uses.

Every recipe is built around three principles:

1. Capsa supplies the context. The connector tells the agent what follow-ups are outstanding, who the customer contact candidates are, and what evidence would close the loop.
2. The user approves before anything leaves. Agents draft; the user reviews; the connected email provider (or other tool) sends only after explicit approval.
3. Completion is recorded only after it actually happened. A draft is not a send. An agent never marks a follow-up done without send evidence or an explicit user confirmation.

Start here

Just connected Capsa? Read start-here.md — what you can do (with example prompts to try), the four-beat loop every task follows (orient → resolve → act → record), and the safety contract the recipes assume.

To put that guidance into your agent, install the capsa-orientation skill — or copy its body into a system prompt, or let an agent fetch it via capsa_discover_playbooks. It's discovery-first, so it keeps working as Capsa ships new capabilities.

Install the skill pack

The skills ship as a Claude Code plugin (this repo is its own marketplace):

/plugin marketplace add capsa-analytics/capsa-agent-cookbook
/plugin install capsa-cookbook@capsa

Or copy any skill's SKILL.md body into another agent — the steps are framework-agnostic. Re-pull after a connector upgrade so the discovery-first guidance stays current.

Machine-readable index

Agents that reach this cookbook via capsa_discover_playbooks can fetch a map instead of scraping this page:

• llms.txt — a curated, link-first overview (llmstxt.org convention).
• index.json — the same map as structured data: capabilities, skills, patterns, tools, and install commands, with stable paths.

Skills — the flagship pack

Installable, self-contained skills. Each is discovery-first and keeps the user in the approval loop:

• capsa-orientation — read first. Discover what's enabled, resolve names to dimensions, stay in the approval loop, and record completion only on evidence.
• proposal-followup-batch — clear a batch of outstanding proposal follow-ups: draft per contact from Capsa context, review, send through a connected email provider, and mark done only after send evidence.
• sensitive-visit-notice — surface upcoming visits flagged sensitive (e.g. chemical application) by division and service type, draft customer pre-notices for review, send only after per-visit approval.
• renewal-portfolio-triage — pull the full renewal book for a scope, tier it with explainable rules (dollars at risk, time pressure, margin flags, payment and satisfaction risk), and present a ranked worklist the user prunes before any deep dive.
• renewal-deep-dive — build a renewal recommendation for one property from renewal drilldowns and property context: price/hours/terms options with cited evidence. The agent proposes; the user decides.

Reference

The comprehensive map of what the connector exposes — capability-led, and kept in sync with the connector's live capsa_describe_capability.

Capabilities

• Follow-up actions — outstanding proposal follow-ups with contact context; evidence-gated completion.
• Upcoming visits — scheduled visits with property and customer-contact context for notice workflows.
• Property context — search properties and pull production, satisfaction, sales, and visit signals; filter a property book by dimension, metric, or date range.
• Renewal opportunities — the renewal book for a window: prior contract baseline, pipeline by status, retention, change tags, plus per-property service comparisons and prior-year performance review.

Patterns — always-on disciplines a skill applies:

• Resolving ambiguous names — work out which dimension a user's term belongs to (is "Tori Nash" a sales rep, an account owner, or a property?) before filtering or reporting.

Required connectors