Meeting Followup Email

Reference these to inform email tone and terminology.

Read EXAMPLES.md for concrete input/output pairs showing correct and incorrect follow-up emails. Study these patterns before reading transcripts.

Resources

Named Sub-Agents

Use these specialized agents by name with the Task tool (subagent_type):

Each agent knows its tools and behavioral constraints. When dispatching:

• Include the account name and key contact names in every prompt
• Specify the output format needed (structured findings with promises, next steps, pain points)
• Tell agents what you already know so they do not duplicate effort

Sub-Agent Delegation Rules

Delegate to sub-agents when:

• Account has 2+ transcripts in Granola or account folder
• User explicitly requests broader context ("check Slack", "look for prior emails")
• Multi-source research is needed (transcripts + Slack + Gmail + account folder)

Read directly in main context:

• Single transcript with a known path (user specified or only one found)
• Org-context files (small, needed before any dispatch)
• EXAMPLES.md (small, needed for email pattern reference)
• Memory search results

ALWAYS:

• Spawn all independent agents in a SINGLE message (parallel execution)
• Create /tmp/meeting-followup/<account-slug>/ before dispatching
• Wait for ALL agents to return before reading output files
• Max 4 agents per dispatch message

NEVER:

• Use run_in_background: true (MCP tools are unavailable in background sub-agents)
• Begin drafting before all agents have returned
• Use placeholder data in the final email

Sub-Agent Output Persistence

Temp directory: /tmp/meeting-followup/<account-slug>/

Reading pattern after agents return:

1. File exists: Read it with the Read tool (primary source)
2. File missing but Task result >500 chars: Use the Task result text as fallback
3. Both missing: Note the data gap, proceed with available data

Workflow

Phase 1: Context Loading

Load MCP tools once at the start. Call ToolSearch('+obsidian'), ToolSearch('+claude-memory'), and ToolSearch('+workspace-mcp draft') in parallel at the beginning of Phase 1. Do NOT call ToolSearch again later for these tools.

Read these files directly in the main context (small files, needed before any agent dispatch):

1. Read org-context files from Obsidian using individual obsidian_get_file_contents calls (do NOT use obsidian_batch_get_file_contents for org-context files; the combined content exceeds token limits and causes overflow errors):
   • Regal/org-context/org-context.md
   • Regal/org-context/product.md
   • Regal/org-context/differentiators.md
   • Regal/org-context/sales-process.md
   • Regal/org-context/use-cases.md
2. Read tools/references/email-links.md from the repo for link resolution (DocSend collateral, developer docs, platform URLs, placeholder rules)
3. Read Regal/org-context/case-studies.md from Obsidian for industry-matched case study URLs (Quick Lookup table)
4. Read EXAMPLES.md for email patterns

If Obsidian MCP is unavailable, apply the filesystem fallback: vault root at OBSIDIAN_VAULT=/Users/nick.yebra/Library/Mobile Documents/iCloud~md~obsidian/Documents/Core Vault. Use Read, Glob, and Bash with this path prefix.

Phase 2: Locate Transcript(s)

Large file handling: If obsidian_get_file_contents or obsidian_batch_get_file_contents returns ANY token limit error, do NOT retry the same call. Instead, retry with offset: 0, limit: 500 lines. For transcripts, reading the first 500 lines is sufficient since enrichment data from get_meeting already covers action items and signals. For org-context files that are too large, skip the oversized file and proceed with the others.

If the user specified a transcript location (e.g., "the transcript is in Regal/Accounts/ABT Electronics"), go there directly. Read the file and skip to Phase 2b.

If no location specified, use a three-step discovery sequence. Run Step A first; if it returns results, skip Steps B and C.

Path naming conventions differ by folder:

• Granola folder: ALWAYS use a hyphenated slug (e.g., "AAA Life" → AAA-Life, "Ascend Learning" → Ascend-Learning). Paths with spaces will 404.
• Accounts folder: Use the original company name with spaces (e.g., Regal/Accounts/Ascend Learning/). Do NOT hyphenate the Accounts folder path.

Step A: Transcript Pipeline (primary, fastest)

Use find_transcript for unified search across SQLite, enriched sidecars, and raw Granola files in a single call:

ToolSearch('+transcript-pipeline')
mcp__transcript-pipeline__find_transcript({ query: "<account name>", days: 7 })

If results are returned: pick the most recent match and call get_meeting(meeting_id) to retrieve the full record including enrichment data (action items, deal signals, risks, participants). Use the enrichment data to inform the email draft: action items become follow-up bullets, deal signals inform tone and urgency, risks flag what to address proactively. Skip Steps B and C and proceed to Phase 2b.

NEVER read the raw Granola file after get_meeting returns enrichment data. The enrichment data is sufficient. Reading the raw file is redundant and frequently hits the 10K token limit (raw transcripts are often 15K+ tokens). This rule applies even if you found the Granola path in Step A results.

If no results, attempt an on-demand Clari fetch (see references/transcript-discovery.md for the full fallback cascade including SF-ID resolution and search_account_calls). If the transcript-pipeline is unavailable (connection error) or the on-demand fetch returns nothing, proceed to Step B.

Step B: Obsidian Granola directory (fallback)

Search Obsidian in this exact order. All three steps are MANDATORY. Do NOT skip the top-level listing.

1. Enriched sidecars Regal/Granola-Enriched/ (ALWAYS check this first). Enriched files mirror the raw Granola directory structure but include action items, deal signals, risks, and participants extracted by the pipeline. Search Regal/Granola-Enriched/{account-slug}/ for account-specific enriched transcripts. If found, prefer these over raw files.
2. Top-level Regal/Granola/ (list as a separate obsidian_list_files_in_dir call). Scan ALL returned filenames for the account name or keywords. Granola frequently places transcripts here with titles that do NOT match the account subfolder pattern (e.g., Room-for-Ascend-Regal-AI-Agent-Demo.md, Ascend-Regal-AI-Agent-Demo.md). Do not assume only Untitled_*.md files live here. Skip files that already have an enriched sidecar in Granola-Enriched/.
3. Account subfolder (Regal/Granola/{account-slug}/): List this last. Skip files that already have enriched sidecars.

Step C: Account folder (supplementary)

• Account folder (Regal/Accounts/{company name}/Meeting Transcripts/): List here for additional transcripts not in Granola, using original spacing.

Recency Validation

After collecting all candidate transcripts from Steps A-C, sort by date (created_at or updated_at). Default to the most recent transcript. If the most recent transcript is more than 48 hours old and the user did not specify which meeting, confirm with the user before drafting: "The most recent transcript I found is from [date]. Is this the meeting you want to follow up on?"

Phase 2b: Dispatch Decision

Route based on what was found:

• Single transcript, user specified path: Read directly in main context (fast path). Skip to Phase 4.
• 2+ transcripts for this account: Dispatch transcript-locator then transcript-analyzer per Phase 3. Do NOT read multiple transcripts directly in the main context; delegate to keep context clean.
• User requests broader context ("check Slack", "look for prior emails", "also check Gmail"): Dispatch additional agents per Phase 3.

Phase 3: Sub-Agent Dispatch

Create the temp directory first: rm -rf /tmp/meeting-followup/<account-slug> && mkdir -p /tmp/meeting-followup/<account-slug>

Dispatch up to 4 agents in a SINGLE message. Each dispatch prompt includes the account name, key contact names, Output Persistence block with /tmp file path, Source URL Guidance, and what you already know (to prevent duplicate effort).

Dispatch prompt templates:

Wave 1: transcript-locator (always dispatched first in multi-transcript path):

Discover all transcript locations for [Account Name]. Key contacts: [names]. Frontmatter tag hints for pre-filtering: folder/[account-slug], person/[contact1-slug], person/[contact2-slug]

Output Persistence: Write discovered transcript paths and metadata to: /tmp/meeting-followup/[account-slug]/transcript-locator.md

If the Write tool fails with "File has not been read yet", Read the target path first, then Write again. If it fails for any other reason, return your FULL findings as text instead (do NOT summarize).

WRITE EARLY: After your first 2-3 successful research calls, IMMEDIATELY write your current findings to the output file. Do NOT wait until all research is complete. Continue researching and UPDATE the file with additional findings. To update: Read the output file first, then use Edit to append or Write to replace. The initial write is your insurance against output loss. A partial file is infinitely better than no file.

For every finding, include the vault-relative path (e.g., Regal/Granola/filename.md).

Locator failure handling (three-state):

1. Success: Locator returns transcript paths. Proceed to Wave 2.
2. Empty: Locator found no transcripts. Log the gap, skip Wave 2, proceed to Phase 4 with available data.
3. Crash: Locator agent failed (no output file, no usable Task result). Log the failure, skip Wave 2, proceed to Phase 4 with available data.

Wave 2: transcript-analyzer (gated on locator success, dispatched with other agents):

Read and extract structured data from ALL transcripts discovered by the locator for [Account Name]. Key contacts: [names]. Locator output: /tmp/meeting-followup/[account-slug]/transcript-locator.md

RECENCY MODE: email-followup Sort all discovered transcripts by date. Full-read the LATEST transcript and write its raw content to the latest-transcript file. For older transcripts, extract focused structured data (promises, next steps, pain points) only.

Output Persistence: Primary output file: /tmp/meeting-followup/[account-slug]/transcript-analyzer.md Latest transcript file: /tmp/meeting-followup/[account-slug]/transcript-analyzer-latest-transcript.md

If the Write tool fails with "File has not been read yet", Read the target path first, then Write again. If it fails for any other reason, return your FULL findings as text instead (do NOT summarize).

WRITE EARLY: After your first 2-3 successful research calls, IMMEDIATELY write your current findings to the output file. Do NOT wait until all research is complete. Continue researching and UPDATE the file with additional findings. To update: Read the output file first, then use Edit to append or Write to replace. The initial write is your insurance against output loss. A partial file is infinitely better than no file.

For every finding, include the vault-relative path (e.g., Regal/Granola/filename.md).

Dispatch transcript-analyzer in the same message as google-workspace-researcher, slack-researcher, and obsidian-analyzer (up to 4 agents total per message).

google-workspace-researcher (when prospect email is known):

Search Gmail for prior email threads with [Contact Name] ([email]). Find the most recent 3-5 email exchanges. Extract: what topics were already covered, what was promised, what questions were asked.

Already known from Phase 2: [Account confirmed: yes/no, transcript found: yes/no with source, enrichment data: available/none, key contacts identified: list names]

Output Persistence: Write complete findings to: /tmp/meeting-followup/[account-slug]/google-workspace-researcher.md After writing, return a 2-3 paragraph summary only. If Write fails, return full findings as text.

If the Write tool fails with "File has not been read yet", Read the target path first, then Write again. If it fails for any other reason, return your FULL findings as text instead (do NOT summarize).

WRITE EARLY: After your first 2-3 successful research calls, IMMEDIATELY write your current findings to the output file. Do NOT wait until all research is complete. Continue researching and UPDATE the file with additional findings. To update: Read the output file first, then use Edit to append or Write to replace. The initial write is your insurance against output loss. A partial file is infinitely better than no file.

slack-researcher (when internal context would help):

Search Slack for recent discussions about [Account Name]. Look for: follow-up commitments, deliverables shared, internal deal context, team discussions about the account. Also search the #ai-demo-requests channel for the account's custom demo agent phone number. Focus on messages from the past 2 weeks.

Already known from Phase 2: [Account confirmed: yes/no, transcript found: yes/no with source, enrichment data: available/none, key contacts identified: list names]

Output Persistence: Write complete findings to: /tmp/meeting-followup/[account-slug]/slack-researcher.md After writing, return a 2-3 paragraph summary only. If Write fails, return full findings as text.

If the Write tool fails with "File has not been read yet", Read the target path first, then Write again. If it fails for any other reason, return your FULL findings as text instead (do NOT summarize).

WRITE EARLY: After your first 2-3 successful research calls, IMMEDIATELY write your current findings to the output file. Do NOT wait until all research is complete. Continue researching and UPDATE the file with additional findings. To update: Read the output file first, then use Edit to append or Write to replace. The initial write is your insurance against output loss. A partial file is infinitely better than no file.

obsidian-analyzer (when account has research docs):

Read files in Regal/Accounts/[Account Name]/. Extract: deal background, key contacts with roles, timeline, recent activity, open action items.

Already known from Phase 2: [Account confirmed: yes/no, transcript found: yes/no with source, enrichment data: available/none, key contacts identified: list names]

Output Persistence: Write complete findings to: /tmp/meeting-followup/[account-slug]/obsidian-analyzer.md After writing, return a 2-3 paragraph summary only. If Write fails, return full findings as text.

If the Write tool fails with "File has not been read yet", Read the target path first, then Write again. If it fails for any other reason, return your FULL findings as text instead (do NOT summarize).

WRITE EARLY: After your first 2-3 successful research calls, IMMEDIATELY write your current findings to the output file. Do NOT wait until all research is complete. Continue researching and UPDATE the file with additional findings. To update: Read the output file first, then use Edit to append or Write to replace. The initial write is your insurance against output loss. A partial file is infinitely better than no file.

Phase 3b: Read Agent Output

After all agents return, read output files from /tmp/meeting-followup/<account-slug>/:

1. Read transcript-analyzer-latest-transcript.md (primary drafting source: full raw content of the most recent meeting)
2. Read transcript-analyzer.md (structured extraction across all meetings: promises, next steps, pain points)
3. Read any additional agent output files that exist

Apply the reading pattern from Sub-Agent Output Persistence: file exists -> read it, file missing but result >500 chars -> use result, both missing -> note gap.

Phase 4: Memory Context

Call search_memories (MCP tool) with the account name and limit: 3. NEVER pass a limit above 3. This surfaces prior observations about communication preferences, past email feedback, and deal context. If unavailable or no results, proceed without. If the result exceeds the token limit and gets persisted to a file, do NOT read the file. Skip memory context entirely and proceed.

Phase 5: Pre-Draft Decision

Answer four questions before writing anything. These determine the entire email.

1. Who is this going to? Prospect, channel partner, internal team, executive? Their role dictates what belongs in the email.
2. What is the ONE thing they need to do next? Book a meeting, send a file, make a decision, nudge someone, approve something? The email exists to get this done.
3. What do they already know? If they were on the call, do not recap it. If they are a partner who brought you the deal, do not sell them on the opportunity. If they already understand the product, do not explain capabilities. Now informed by google-workspace-researcher's prior email threads when available.
4. What work have I done since the call? Customized a demo, updated configurations, built an agent, found a document, prepared materials? List every concrete deliverable with specifics (names, numbers, links). These deliverables are often the core of the email: they prove responsiveness and move the deal forward. Now informed by slack-researcher's internal discussions when available.

Then apply the right approach:

• Prospect who was on the call: Reinforce their confidence that you are the right solution. Anchor on their workflow and pain points (not your features). Deliver any promised assets. Make the next step feel inevitable, not optional. For pre-demo follow-ups: ask for sample call scripts, outreach workflows, or process documents they want you to reference during the demo. This shows you will build around their workflow, not run a generic demo.
• Channel partner: They already believe in you. Just confirm you can do what the client needs, tell them what you are building, and ask them to do the one thing only they can do (usually nudge the client or book a meeting).
• Internal team: Orient them on the deal, transfer context, assign actions. Use the handoff structure from writing-style.
• Executive or decision-maker not on the call: Lead with the business outcome, not the technical discussion. Keep it to 2-3 sentences max.

Phase 6: Draft Email

Follow all rules in the Email Guidelines section below. Apply the writing-style skill rules including pyramid principle document structure.

Phase 7: Deliver

Present ONLY the clean email in a clear To / CC / Subject / Body format that is easy to copy-paste. Do NOT include the pre-draft decision, reasoning notes, or internal analysis in the output. The user should see nothing except the finished email. Include a short, specific subject line (e.g., "Next steps from our call" or "Follow-up: integration timeline"). If multiple distinct recipients need different action items, draft separate emails.

Gmail draft creation: When the user confirms the email or asks to draft it in Gmail, ALWAYS use body_format: "html" with draft_gmail_message. Plain text (body_format: "plain") strips all formatting: no bold, no clickable links, no bullet structure. Convert the email to HTML before sending: use <b> for bold headers, <ul>/<li> for bullets, <a href="..."> for links, <br> for line breaks, and <p> for paragraphs. Always set include_signature: true.

Email Guidelines

Structure: Choose the right pattern based on Question 4 (deliverables):

• Deliverables-first (when you have work to show): Pleasantry → framing line ("Here's what we discussed moving forward:") → bullet list of deliverables with specifics → next step ask embedded naturally with named people → optional soft secondary request → close. This is the default for post-demo and post-workshop follow-ups.
• Ask-first (no deliverables, just scheduling or decisions): The first sentence states what happens next or what you need from them. Not what was discussed. Not what Regal can do. What happens next. Everything after that is evidence or context supporting that first sentence.

NEVER lead with a recap of what was discussed. NEVER lead with what Regal can do. Lead with what you have built or what happens next.

Tone: Short and casual. Action-oriented, not discussion-focused.

Subject lines: Keep under 8 words. Lead with the purpose, not the meeting name.

• Good: "Next steps from our onboarding call"
• Good: "Follow-up: integration timeline"
• Bad: "Meeting Follow-Up 2/10/2026"

Link resolution: Resolve all links from tools/references/email-links.md and case-studies.md. NEVER use a placeholder for a link that exists in either file.

• DocSend collateral (decks, one pagers, legal/security docs): Use URLs from email-links.md. Match the prospect's industry to vertical one-pagers when available.
• Developer docs: Use URLs from email-links.md. Match the "When to include" column to the conversation topic.
• Platform links (self-serve demo, Reggie demo number): Use URLs from email-links.md. Common transcript aliases: "sandbox", "sandbox access", "platform access", "platform tour" all resolve to the self-serve platform demo URL (https://www.regal.ai/privatetour).
• Case studies: Match the prospect's industry to the Quick Lookup table in case-studies.md. Use the actual Link column URL. Only include public or sales-approved entries (never internal).
• Custom demo numbers: Check the #ai-demo-requests Slack channel for the account's demo agent phone number before using a placeholder. Dispatch slack-researcher with an explicit instruction to search this channel. Only use [Insert DEMO NUMBER] if Slack search returns nothing.
• Placeholder only when unknown: Account-specific recordings, Google Drive files, dates for scheduled meetings. Use clear format: [Insert LINK to recording], [Insert DATE].

Handling MY promises (things I said I would do):

• Quick tasks (<5 min like finding a document, looking up info, rescheduling): Assume already done, use placeholders
  • Example: "I rescheduled our next meeting to [Insert DATE]"
  • Example: "Here's the DPA: [Attach DPA]"
• Longer/important tasks: State what you did with specifics. Show the work, do not just claim it.
  • Good: "Updated the demo orders to use Ed, Phil, Jon, Mike, & Ricky for more personalized testing, along with the ABT signature phrasing you mentioned"
  • Good: "I sent over a request to our engineering team to build the custom demo"
  • Bad: "I'll follow up with our engineering team on the demo"
  • Bad: "We updated the demo" (too vague — name what changed and for whom)

Handling THEIR promises (things others said they would do):

• Push toward action with questions instead of statements
  • Instead of: "Amanda needs to send the budget figures"
  • Write: "Amanda, when do you think you'll be able to send over the budget figures?"

Proactive collateral: Don't just include links discussed on the call. Proactively include relevant resources using email-links.md and case-studies.md:

• Case studies: Match the prospect's industry to the Quick Lookup table. Select 1-2 case studies where Shareable is public or sales-approved. Use the actual URL, One-liner for context, and Top Metric for impact.
• Vertical one-pagers: If a matching vertical one-pager exists in email-links.md (Insurance, Medicare, Healthcare, Education, Loan Servicing), include it.
• Platform demo: Include the self-serve demo URL from email-links.md when a demo was discussed or offered.
• Developer docs: Include relevant doc pages from email-links.md when technical next steps were discussed (agent building, integrations, API access). Include these in an "In the meantime" or "Additional resources" section. NEVER use a placeholder for a link that exists in either file.

Formatting rules:

• Never quote the transcript directly in the email
• Keep it brief. For prose-style emails, 3-5 sentences. For deliverables-first emails, use a short header + bullet list + 1-2 closing sentences. Bullets can be longer than sentences but should each contain one idea.
• Always include proof points, demo assets, or links that were discussed on the call (demo numbers, documentation links, recordings). If you do not know the exact link, use a placeholder.
• Never use double conditionals ("if you'd like... if so..."). Use assumptive close language for agreed-upon next steps. "Whenever you're ready" not "If you'd like to move forward, if so we'd..."
• If something was confirmed as doable on the call, state it as a capability. Do not hedge it, do not say "I'll check on that," do not reintroduce uncertainty.
• End with a clear next step or a specific question

Banned words and phrases (always enforced):

• Words: delve, foster, underscore, pivotal, showcase, landscape, robust, leverage (as verb), utilize, streamline, spearhead, bolster, cutting-edge, synergy, holistic, wedge
• Padding phrases: "it's worth noting," "importantly," "at the end of the day," "it's important to remember," "in today's [anything]," "when it comes to"
• Hedging words: should, might, could, probably, maybe
• Grandiose claims: paradigm shift, game changer, revolutionary, unprecedented, transformative

Multiple recipients: If attendees have different action items, ask the user whether to draft one group email or separate emails. Default to one email with people's names inline next to their items.

After Delivery

After presenting the email, ALWAYS include this line at the end of your response:

Run /track-tasks to extract and track action items from this meeting.

This is a required step, not optional. Do not skip it even if the email is short or the meeting had few action items.