---
name: curcuma-email
description: |
  Inbox intelligence for coding agents — lifecycle CRM, subject/CTA copy,
  competitive research, production/Outlook patterns, plus design and improve.
  Ground every answer in Curcuma MCP tools over ~19.8k production emails.
  Routes to the right job skill instead of guessing from training data.
trigger: When asked to design, write, improve, research, plan lifecycle/CRM email, subject lines, CTAs, competitive email research, or Outlook/production email HTML.
---

# Curcuma Email (Router)

Curcuma is **not only design taste**. It is a production-email library your
agent queries for lifecycle, copy, competitive intel, production engineering,
and HTML craft.

Do not reimplement a mode here. Read the selected skill’s `SKILL.md` and follow it.

## Routing (choose exactly one)

| User intent | Skill |
|---|---|
| Lifecycle / CRM moment or sequence (cart, welcome series, winback, trial, cancel, shipping/receipt…) | `skills/curcuma-lifecycle/SKILL.md` |
| Subject lines, preheaders, CTA/button copy, offer language | `skills/curcuma-subject-cta/SKILL.md` |
| Competitive / vertical / category research (“how do brands…”, benchmarks) | `skills/curcuma-competitive/SKILL.md` |
| Outlook, MSO, bulletproof buttons, responsive, dark mode, footer compliance | `skills/curcuma-production/SKILL.md` |
| Write a **NEW** full HTML email from a brief | `skills/curcuma-email-design/SKILL.md` |
| Improve / fix **existing** email HTML | `skills/curcuma-email-improve/SKILL.md` |
| General pattern/type research (not competitive framing) | `skills/curcuma-email-research/SKILL.md` |

### Disambiguation

- “Write me a welcome email” (wants HTML) → **design**, not lifecycle
- “What should our post-signup email sequence be?” → **lifecycle**
- “Give me 10 subject lines for cart abandonment” → **subject-cta**
- “How do beauty brands do abandoned cart?” → **competitive**
- “This breaks in Outlook” / “bulletproof button” → **production**
- “Fix this HTML” (paste) → **improve**
- Bare `/curcuma-email` → list the table above and ask which job fits

## Prerequisites

Curcuma MCP must be configured. Probe with `search_emails`.

**Hosted:** mint at `POST https://curcuma.sud.fyi/api/mcp/install-token`
and connect streamable HTTP MCP at `https://curcuma.sud.fyi/mcp`
with `Authorization: Bearer <token>`.

If tools are missing, tell the user to open the landing and mint a key, or re-run
the installer.

## Search discipline (every skill)

- Always run at least one real search for the user’s actual type/moment
- Prefer 2–4 word queries; on weak/none coverage, simplify or `list_categories`
- Never repeat an identical query; page with `offset` / `next_offset`
- `max_per_company=1` for diversity unless studying one brand
- Prefer `get_email_summary` over `get_email_html`
- Never invent prevalence — only quote `get_pattern` or in-sample counts
- Never invent brand journey timelines or conversion metrics

## Grounding (required — read `https://curcuma.sud.fyi/.well-known/agent-skills/curcuma/GROUNDING.md`)

**Corpus before imagination.** If tools are available, call them before any
email HTML, demo/proof pair, subject list, lifecycle plan, or competitive claim.

Hard fails (do not ship):

| Fail | Do instead |
|---|---|
| Invented “proof” demos with fake left/right arms | Real eval output + real corpus slugs / skeletons |
| Dashed `{{IMG_HERO}}` / placehold.co as “finished” | Real URLs or honest tokens + `needs_images[]` |
| Therapy-speak AI copy (“no guilt”, “hope this finds you well”) | Tone from `visible_text` / real CTAs in summaries |
| Generic indigo SaaS chrome for every brand | `color_palette` + layout from matched summaries |
| “Inspired by real emails” with no slugs | Cite every slug you used |
| Plain text passed off as production email HTML | Table-based HTML; plain text only if user asked |
| Silent ungrounded write when MCP is down | Label **ungrounded** or fix MCP / mint key |

Minimum write loop: `search_emails` → `get_email_summary` (×2–3) →
`get_pattern` when structure matters → write → report slugs.

## Tool reference

| Tool | Use |
|---|---|
| `search_emails` | Find references by keyword/category/company |
| `get_email_summary` | ~3KB study brief (subject, preheader, CTAs, structure) |
| `get_pattern` | Corpus prevalence + production snippets |
| `list_categories` | 204 categories |
| `get_skeleton` | De-branded chassis when writing HTML |
| `get_email` / `get_email_html` | Metadata / raw HTML when truly needed |

## Anti-patterns

- DON'T invent emails from training memory when Curcuma is available
- DON'T dump 60KB HTML to “study” when summary exists
- DON'T claim open rates, lift, or ordered multi-brand journeys without data
- DON'T use fake image hosts (placehold.co, etc.) in final HTML
- DON'T invent before/after demos — use corpus + eval artifacts
- DON'T ship therapy-speak or emoji-spam heroes as “taste”
- DON'T omit slug citations on grounded deliverables