PREREQUISITE — Identify the sender (who is running this skill)

The user has invoked the email follow-up v2 skill. The argument (if provided) is: $ARGUMENTS

---

Guard check: If $ARGUMENTS is empty or the literal string $ARGUMENTS, ask the user: "Which meeting or contact should I draft a follow-up for?" — then wait for their response and use that as the search term throughout.

---

PREREQUISITE — Identify the sender (who is running this skill)

Do this BEFORE fetching any Fireflies or HubSpot data, so meeting-host names and deal-owner names cannot contaminate the identity. Resolve the sender's name + email and store it as SENDER — this is the person who will sign the draft in Step 6.

Priority (stop at the first that succeeds):

1. The authenticated Gmail user's profile (use the Gmail MCP's profile / get_me call if available) → take name and email
2. Else, the display name on the From header of the user's 3–5 most recent sent emails
3. Else, the local-part of the authenticated Gmail address, title-cased (e.g. [email protected] → "John Smith")
4. Else, stop and ask the user directly: "I couldn't identify your name from Gmail. What name should I sign the email with?" — then wait for their reply and use that verbatim

Hard rules — never break these, no matter how tempting the signal looks:

• SENDER is NOT the Fireflies meeting host, organizer, or recorder
• SENDER is NOT any internal participant name found in the Fireflies transcript
• SENDER is NOT the HubSpot deal or contact owner (hubspot_owner_id / search_owners result)
• SENDER is NOT inferred from any name that appears in the transcript, engagement history, or HubSpot record

SENDER only comes from the sources in the priority list above. If the priority list fails, you must ask the user — never guess.

---

Step 1 — Fetch Fireflies transcripts (up to last 3, chronological)

Search Fireflies for recent meeting transcripts matching $ARGUMENTS:

• Use fireflies_search / fireflies_list_transcripts to find matching transcripts
• Select up to the 3 most recent transcripts within the last 90 days that match the argument (by title, participant, or company)
• For each selected transcript, call fireflies_get_transcript and extract: meeting date, participants (name + email), summary, structured action_items, key topics
• Order the transcripts oldest → newest before using them downstream so the draft can reflect progression ("since our first call… on our last call…")
• If only one transcript matches, proceed with just that one
• If no transcript matches: say "No Fireflies transcript found for '$ARGUMENTS'. Please paste your meeting notes below and I'll use those instead." — then wait

Also detect the transcript language (from Fireflies metadata if available, else infer from the transcript text). Record it for Step 7.

---

Step 2 — Resolve HubSpot contacts by email

From the most recent transcript, collect external attendee emails (exclude the rep's own domain).

For each attendee email:

• Use search_crm_objects with objectType: "Contact" and filter on email equals the attendee's address — this returns a unique contact (no name ambiguity)
• Fetch: firstname, lastname, email, jobtitle, hubspot_owner_id, associated company
• Resolve hubspot_owner_id to a name via search_owners

If a transcript attendee has no email, fall back to a name search for that one attendee only.

Also look up the associated deal:

• Use search_crm_objects with objectType: "Object" on the company or dealname derived from $ARGUMENTS / the contact's company
• Fetch: dealname, name_of_deal_stage, pipeline, hubspot_owner_id

If neither a Contact nor Deal is found AND no Fireflies transcript was found in Step 1: stop and say "I couldn't find a Fireflies transcript or a HubSpot record for '$ARGUMENTS'. Please paste your meeting notes or provide a HubSpot contact/deal name."

If HubSpot has no matches but Fireflies does, proceed — the recipient emails come from the transcript.

---

Step 3 — Pull HubSpot engagement history (cost-bounded)

Fetch engagements on the deal / primary contact with these filters to keep token cost low:

• Recency: last 90 days only
• Count cap: top 10 most recent engagements
• Type priority: prefer Note and Call engagements first; include Email only if there's room under the cap
• Email truncation: for email engagements, keep subject + first ~200 characters of body only
• Pre-summarize immediately: compress all fetched engagements into a 5-bullet relationship memo (objections raised, stakeholders mentioned, pricing discussed, commitments made, current status). Use only the memo in Step 6, not the raw engagement data — this keeps edit iterations cheap.

If no engagements exist (new relationship), skip silently.

---

Step 4 — Check for prior Gmail thread

Search Gmail for the most recent thread with the primary recipient's email (or their company domain), limited to the last 30 days:

• If a thread is found: capture the subject line and the last 1–2 messages (truncated to ~300 chars each). Use this as context so the draft references the ongoing conversation naturally.
• If no thread: proceed without prior-thread context.

This is context-only — we are not pushing a reply draft to Gmail. Output stays in Claude.

---

Step 5 — Sample rep's writing voice (last 5 sent emails)

Read the user's last 5 sent emails from their Gmail Sent folder (or equivalent). Extract lightweight style signals:

• Greeting style (e.g. "Hi [name]," vs "Hey [name]," vs "Hello [name],")
• Average sentence length and paragraph density
• Sign-off phrase (e.g. "Best," / "Cheers," / "Thanks,")
• Formality level and contraction usage
• Any recurring closing phrases

Store these as a short voice profile for Step 6. This step only affects tone/phrasing — it does not determine the sender's name. The sender was already locked in the PREREQUISITE step as SENDER. Do not overwrite SENDER here.

If the user has no prior sent emails: skip silently and fall back to the professional baseline only. No prompt.

---

Step 6 — Generate the email draft

Write a follow-up email using the data gathered in Steps 1–5.

Tone:

• Baseline (always applied): professional, concise, no slang, no emojis, no over-familiarity
• Voice overlay: layer the rep's voice profile from Step 5 on top of the baseline. If no profile exists, use baseline only.
• Do not vary tone by deal stage.

Structure:

1. Opening — reference the meeting(s) specifically. If multiple transcripts were found in Step 1, reflect the progression across them.
2. Summary — 2–3 bullets recapping what was discussed (synthesize across transcripts if multiple). Draw on the engagement-history memo (Step 3) for relevant prior context.
3. Next steps — use only the structured action_items from Fireflies. If none exist, write: - No action items captured — add manually. Do not infer action items from transcript prose.
4. Sign-off — use the sign-off phrase from the voice profile (Step 5), followed by SENDER.name from the PREREQUISITE step. Do NOT use any name from the Fireflies transcript (host, organizer, recorder, or any participant). Do NOT use the HubSpot deal/contact owner. Do NOT use any name that appears in the engagement history. SENDER.name is the only valid source for the sign-off name. It will always be resolved, because the PREREQUISITE step either found it in Gmail or asked the user directly.

Subject line: Follow-up: [Company Name] x Pliant

• If no company/contact name is available, use the raw $ARGUMENTS value.

Recipients:

• To: primary contact (the main attendee from the most recent transcript, resolved in Step 2)
• CC: other external attendees from the transcript who have emails

---

Step 7 — Language

Draft the email in the language detected in Step 1 (Fireflies transcript language). If detection failed, use the language the rep has been using in the current Claude session. Apply the voice overlay from Step 5 regardless of language (the stylistic signals translate).

---

Step 8 — Render the full email in Claude

Display the email to the user in this format:

**Context used:**
- Fireflies: [N transcript(s): title + date each, oldest → newest] / [Manual notes provided]
- HubSpot: [Contact name(s), job title(s), company] / [Deal name, stage] / [Not found]
- Engagement memo: [5 bullets from Step 3, or "none"]
- Prior Gmail thread: [subject + date] / [none]
- Voice profile: [brief — e.g. "Hi + short sentences + 'Cheers,' sign-off"] / [baseline only — no prior sent emails]
- Signing as: `SENDER.name <SENDER.email>` — resolved in the PREREQUISITE step
- Language: [detected language]

---

**Draft:**

To: [primary email]
Cc: [other attendee emails, or "none"]
Subject: [subject line]

[email body]

Then say: "Let me know if you'd like any changes, or copy this into Gmail when you're happy with it."

If the user requests edits, apply them and re-render. Repeat until approved. Do not push to Gmail — v2 outputs the email in Claude only; the rep copies it over themselves.