email-find

Email Find

A playbook for finding a specific email through the Gmail MCP (claude.ai Gmail) when the obvious search query returns nothing. Optimized for the case where the user describes an email and the agent must not give up easily.

Usage: /email-find <description>

ARGUMENTS: $ARGUMENTS

Behavioral rules

1. If the user is confident an email exists, the email exists. Do not respond with "I can't find it" after one or two failed queries. The default failure mode is the agent's query, not the user's memory.

2. Broad before narrow. Always run at least one query that scans the full inbox by recency before narrowing on subject or sender keywords. A reply about Topic-X can land on a thread whose subject names a different topic, with zero topic-X tokens in the subject — a keyword filter throws those threads away.

3. Expand sender to partners, not just the named entity. When the user mentions a customer / vendor / collaborator by name, also search for known associated email domains. Operational replies often come from data-ops partners, contractors, or assistants — not the named principal contact.

4. Re-fetch threads in full. get_thread with MINIMAL format returns metadata + snippets for every message but omits the message bodies. If you've identified the candidate thread and need to read what was actually written, fetch it again with FULL_CONTENT. The search-result preview in the UI may also truncate long threads; FULL_CONTENT exposes everything.

5. Show the search trail. End every "found it" or "still hunting" reply with the list of queries you tried, so the user can see what worked and what didn't.

Workflow

In the queries below, me is Gmail's reserved keyword for the authenticated user's primary address — works for everyone regardless of which account is connected.

Phase 1 — Broad scan

Run one broad query first to anchor on what's actually in the inbox in the relevant time window:

search_threads query="(to:me OR from:me) newer_than:Nd" pageSize=15

Where N covers the window the user cited (default 2; cite-driven). The (to:me OR from:me) form covers both received and sent mail — stubborn lookups are sometimes for a message the user sent and can't refind. Look at the actual returned threads — note senders, subjects, dates. Often the email is already in the top 10 results, just with a subject you wouldn't have guessed.

Phase 2 — Expand sender domain

If Phase 1 didn't surface it, run one query per partner domain the user may have meant. Look up known partner domains for the named entity in ## Per-user partner-domain memory below. For each, format:

search_threads query="from:DOMAIN OR from:NAMED-ADDRESS" pageSize=10

DOMAIN here is the bare domain (e.g. acmecorp.com), not a wildcard like *@acmecorp.com — Gmail's from: operator matches any address at the bare domain but does not support *@ wildcards on the user portion. If the memory file stores domains in *@domain form for readability, strip the *@ prefix when building the query, otherwise Phase 2 silently no-ops.

If no partner-domain file exists yet, skip this phase and proceed to Phase 3. When Phases 3–4 later surface an email from an unexpected domain, auto-record the mapping per ## Per-user partner-domain memory below.

Phase 3 — Walk threads by participant

If Phase 2 didn't surface it, list threads where the named contact appears anywhere on the message (To/CC/BCC), not just as sender:

search_threads query="DOMAIN OR NAMED-ADDRESS"

Gmail q-syntax treats a bare email as an any-position match — covers from:, to:, and cc:.

Phase 4 — Re-walk identified threads

If you've spotted a candidate thread (e.g. one whose subject matches a related topic, or whose participants overlap), fetch it again with messageFormat: "FULL_CONTENT" and scan ALL messages, not just the snippet on the search hit. The actual answer is often inside a thread you've already seen but truncated.

Phase 5 — Ask only after Phases 1–4

If all four phases came up empty, then ask the user for one of:

• Exact sender email
• Exact subject line or a short snippet
• Approximate timestamp (UTC or local — both work)

When you ask, include the actual list of senders + subjects you saw, so the user can spot the email and tell you which one.

Per-user partner-domain memory

Named-entity → partner-domain mappings live in the user's memory directory, not in this skill.

File location

Resolve SUTANDO_MEMORY_DIR (default: $CLAUDE_CONFIG_DIR/projects/<project-id>/memory/). Look for reference_partner_domains.md, partner_domains*.md, or a ## Partner domains section in user_profile.md. If none exist, the skill remains fully functional — Phase 2 just no-ops.

Auto-learn

When Phases 3–4 surface an email from a domain the user didn't name for that entity, append the mapping to the partner-domains file silently. Do not ask for confirmation; do not narrate the save in the reply. If the file doesn't exist, create it on first discovery. The cost of an unhelpful row is one extra query in a future fanout; the cost of asking is friction every time.

File format

---
name: partner-domains
description: Named entities → associated email domains. Auto-maintained by /email-find.
metadata:
  type: reference
---

| Named entity | Associated email domains |
|---|---|
| Acme Corp | `acmecorp.com`, `acme-data-ops.com` |
| Foo Foundation | `foo.org`, `[email protected]` |

Store bare domains (e.g. acmecorp.com) — Phase 2 uses them verbatim in from:DOMAIN. Specific addresses (e.g. [email protected]) are fine alongside bare domains. No timestamp bookkeeping: a stale row only costs one extra Phase-2 query in fanout, and hand-maintained first_seen / last_useful / pruned_at would be applied inconsistently turn-to-turn — heavy protocol for low payoff. Match whatever frontmatter convention the user already uses elsewhere in their memory dir.

Subject-mismatch heuristic (no subject filtering in Phases 1–3)

A reply about Topic-X frequently rides on an existing operational thread whose subject is about something entirely different. The most common cases:

• A topic that started as a complaint/incident keeps using the incident's subject for all subsequent replies, even months later.
• A customer's data-ops team replies to whatever was the first email in the relationship, ignoring topic shifts.
• A forwarded thread (Fwd: Fwd: ...) carries the original subject forever.

Implication: never subject-filter on the named entity in Phases 1–3. Subject keywords go in Phase 5 only, after the user provides them. Trust sender / recipient / date scoping; let the subjects be whatever Gmail kept on the thread.

Reporting

After running the workflow, reply with:

1. The candidate email's sender, subject, timestamp, and attachment list
2. Which phase found it (1–4)
3. What queries were run in each phase (one line each)
4. If Phase 4 was used, which thread was re-walked

If nothing was found after Phase 4, reply with:

1. The 10 most recent emails in the cited time window (sender + subject + timestamp)
2. All queries tried
3. A direct request for clarification on one of: sender, subject, timestamp

Don't

• Conclude "not found" before completing Phases 1–4.
• Show clarification questions before showing the broad-scan list.
• Hardcode partner mappings in this skill — they live in user memory.