Create Proposal

Delegate to sub-agents: All prospect research (5 agents dispatched in parallel in Phase 2, Step 2a). Each agent writes findings to /tmp/create-proposal/<account-slug>/<agent>.md.

NEVER in main context: Batch-read Salesforce, Obsidian account files, or transcripts directly (delegate to agents). Call 3+ MCP research tools directly (spawn sub-agents instead). Dispatch agents with run_in_background: true (MCP tools unavailable in background).

Named Sub-Agents

Sub-Agent Output Persistence

Read tools/references/output-persistence.md for the pattern and fallback protocol.

Pre-dispatch directory creation:

rm -rf /tmp/create-proposal/<account-slug> && mkdir -p /tmp/create-proposal/<account-slug>

Agent-to-file mapping:

Persistence block (include verbatim in every dispatch prompt, replacing placeholders):

Output Persistence:
Write your complete, detailed findings to: /tmp/create-proposal/<account-slug>/<agent-name>.md
Format as markdown with all data, evidence, URLs, and citations.
After writing the file, return a 2-3 paragraph summary of key findings only.
The file is the primary deliverable. The returned text is a backup summary.
If the Write tool fails, return your full findings as text instead.

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.
The initial write is your insurance against output loss. A partial file is
infinitely better than no file.

Three-tier fallback after dispatch:

Phase 1: INTAKE (research-first, confirm-second)

Pre-populate all intake fields from existing CRM and vault data before asking the user to confirm. NEVER present blank defaults when real data is available.

Extract the prospect name from $ARGUMENTS. If no prospect name provided, ask:

"Which prospect should I create the proposal for?"

Step 1: Quick Data Pull (3 queries, main context)

Run these in parallel (single message) to pre-fill the intake table:

Salesforce query via mcp__Salesforce-Search__salesforce_query_records:

SELECT Id, Name, Industry,
  (SELECT Id, Name, StageName, Amount, AI_Agent_Use_Case__c, CloseDate
   FROM Opportunities WHERE IsClosed = false ORDER BY CreatedDate DESC LIMIT 1),
  (SELECT Id, FirstName, LastName, Title, Email
   FROM Contacts WHERE Primary_Contact__c = true LIMIT 1)
FROM Account
WHERE Name LIKE '%<prospect>%'
LIMIT 1

If the nested query fails (some orgs restrict subqueries), run two separate queries:

1. Account by name
2. SELECT Id, Name, StageName, Amount, AI_Agent_Use_Case__c FROM Opportunity WHERE Account.Name LIKE '%<prospect>%' AND IsClosed = false ORDER BY CreatedDate DESC LIMIT 1

Obsidian account folder via mcp__obsidian__obsidian_list_files_in_dir for Regal/Accounts/<prospect>/. If the folder exists, read ALL files in recency-priority order:

Priority 1: Meeting Transcripts (latest to oldest). List Meeting Transcripts/ subfolder, sort by date (filename or modified date), read newest first. These are the freshest signal and the primary source for: volumes discussed, AI minute estimates, call duration benchmarks, telephony mentions (SIP/VoIP), contact names and titles, use case details, deal-specific data. Stop reading older transcripts once all intake fields are populated.

Priority 2: Other account docs (latest to oldest). Read remaining files in the account folder (deck outlines, meeting prep docs, demo agent prompts, status files, etc.) sorted newest-first. These fill gaps the transcripts didn't cover: industry context, competitive landscape, pain points, internal Regal notes.

Priority 3: Context.md (base layer). Read last. Context.md is a synthesized brief that may contain outdated information. It serves as fallback for any field still missing after Priorities 1-2. When a transcript contradicts Context.md, the transcript wins.

Key principle: Later information supersedes earlier information. A volume number from last week's transcript overrides a volume number from Context.md written two months ago.

If Obsidian MCP is unavailable, fall back to filesystem at $VAULT_ROOT/Regal/Accounts/<prospect>/.

Step 2: Pre-Fill Intake Fields

Map discovered data to intake fields. Priority: latest transcript > older transcripts > other account docs > Context.md > SF fields > smart defaults. First non-null wins at each priority level.

Fixed values (never vary, omit from editable table): Deal structure = POC, Initial term = 90 days, Minimum spend = $25,000/quarter, Support tier = Base.

Step 3: Present Pre-Filled Table for Confirmation

Present the pre-filled table with source citations so the user knows where each value came from:

Proposal Intake for [Prospect] (pre-filled from SF + Obsidian + Transcripts):

| Field                  | Value              | Source                          |
|------------------------|--------------------|---------------------------------|
| Industry               | Healthcare         | SF: Account.Industry            |
| Use case               | Inbound            | SF: AI_Agent_Use_Case__c        |
| Key contact            | Jane Doe, VP CX    | SF: Primary Contact             |
| Monthly AI minutes     | 50,000             | Transcript: "50K minutes/month" |
| Telephony type         | VoIP               | Default (no data found)         |
| Monthly call volume    | 25,000             | Transcript: volume discussion   |
| Avg call duration      | 5 min              | Default (inbound use case)      |

Fixed values (standard for all proposals):
- Deal structure: POC
- Initial term: 90 days
- Minimum spend: $25,000/quarter
- Support tier: Base

Confirm these values or specify overrides (e.g., "change monthly AI minutes to 30,000"):

Wait for user confirmation before proceeding to Phase 2. Accept overrides to any field.

Phase 2: RESEARCH

Step 2a: Sub-Agent Dispatch

Run rm -rf /tmp/create-proposal/<account-slug> && mkdir -p /tmp/create-proposal/<account-slug> via Bash before dispatching.

Dispatch in THREE waves. Wave 1 runs 4 agents (obsidian-locator, salesforce-researcher, transcript-locator, company-researcher). Wave 2 runs 2-3 agents (obsidian-analyzer with locator paths, transcript-analyzer gated on locator results, edgar-researcher).

After Wave 1 returns, read obsidian-locator output. Pass discovered file paths to obsidian-analyzer's dispatch prompt. If obsidian-locator found no files in the target directories, skip obsidian-analyzer for this account. Note [NO OBSIDIAN DATA: <account>] and proceed with remaining agents. If obsidian-locator fails entirely (crash, timeout, or tool_uses: 0), dispatch obsidian-analyzer without path pre-filtering (graceful degradation).

Wave 1 (dispatch 4 agents in a SINGLE message). Each agent receives the account name, context from intake, the output persistence block, and agent-specific extraction targets:

• obsidian-locator: "List all files in Regal/Accounts/[PROSPECT]/. Identify Context.md, research docs, prior proposals, meeting transcripts in the account folder, pricing discussions. Return file paths, types, and modification dates."

• salesforce-researcher: "Find Account, Opportunities, Contacts, and recent Activities for [prospect]. Extract: deal stage, ACV, contract terms discussed, estimated volumes (monthly AI minutes, call volume), any pricing or discount mentions in Activity notes. Also query AI_Agent_Use_Case__c on the Opportunity."

• transcript-locator: "Find transcripts mentioning [prospect]. Return transcript file paths, dates, and brief descriptions. Write results to /tmp/create-proposal/<account-slug>/transcript-locator.md."

• company-researcher: "Research [prospect] company background. Find: industry vertical, business model, contact center scale (locations, agent count, call volume if public), technology stack, recent news, strategic priorities. ALSO: find a transparent PNG or SVG of the company logo suitable for a presentation cover slide. Search the company website, press kit, media/brand assets page, or Wikipedia. Return the direct image URL."

After Wave 1 returns: ls /tmp/create-proposal/<account-slug>/ then read each file using the three-tier fallback protocol.

Wave 2 (dispatch 2-3 agents in a SINGLE message, after Wave 1 returns):

• obsidian-analyzer: "Read these specific files discovered by obsidian-locator: [LOCATOR_OUTPUT_PATHS]. Extract: contacts with titles, volume data discussed, pricing history, use case details, competitive mentions, prior proposal terms. Prioritize Context.md and research docs over raw transcripts. WRITE findings after reading the first 2-3 files."

• transcript-analyzer: Gated on transcript-locator results. Apply three-state logic:

  • Success (locator returned transcript paths): dispatch transcript-analyzer with the discovered paths. Prompt: "Read these transcripts: [LOCATOR_PATHS]. Extract: specific volume numbers discussed (monthly calls, AI minutes, agent counts), pricing expectations, objections, use case requirements, decision criteria, timeline commitments, and any competitive mentions. Write findings to /tmp/create-proposal/<account-slug>/transcript-analyzer.md."
  • Empty (locator completed but found no transcripts): skip transcript-analyzer. Note [NO TRANSCRIPT DATA] and proceed.
  • Crash (locator failed, timed out, or tool_uses: 0): skip transcript-analyzer. Note [TRANSCRIPT LOCATOR FAILED] and proceed.

• edgar-researcher: "Research [prospect]. Return structured SEC filing data: revenue, growth rate, margins, business segments, strategic priorities from 10-K/earnings calls. If private company, return 'Private company' immediately."

After Wave 2 returns: read output files using the three-tier fallback protocol.

Step 2b: Org-Context Reads

Read these Obsidian org-context files for positioning content. Try Obsidian MCP obsidian_get_file_contents first. If Obsidian MCP is unavailable (connection refused), fall back to reading directly from the filesystem at /Users/nick.yebra/Library/Mobile Documents/iCloud~md~obsidian/Documents/Core Vault/Regal/org-context/.

• Regal/org-context/use-cases.md
• Regal/org-context/case-studies.md
• Regal/org-context/differentiators.md
• Regal/org-context/product.md
• Regal/org-context/sales-process.md (Stage 3 exit criteria for Close slide)

NEVER use obsidian_batch_get_file_contents for these files. These 5 files total 100K+ characters combined and will overflow the batch call's token limit, triggering the context-trap "read in sequential chunks" error. Call obsidian_get_file_contents for each file individually in a single parallel message (5 calls, 1 per file).

Step 2c: ROI Engine Pipeline

Run the ROI pipeline inline (not via nested /business-case dispatch):

1. Detect industry and use case from agent outputs. Check Salesforce AI_Agent_Use_Case__c first, then transcript mentions, then company context. Options: General, Healthcare, Education, Retail. Use case options: Inbound Triage, Inbound Lead Verification, Outbound, Operational.

2. Load matching benchmark YAML: tools/skills/business-case/references/benchmarks/{use_case}/{industry}.yaml

3. Prepare input JSON mapping intake fields to engine keys:

   Intake Field	Engine Key
   Estimated monthly AI minutes	monthly_inbound_volume (inbound triage), monthly_leads (lead verification), monthly_outbound_attempts (outbound), monthly_transactions (operational)
   Estimated monthly call volume	monthly_call_volume
   Use case	use_case
   Industry	industry

4. Run roi_engine.py (Bash tool, from repo root):

   source .venv/bin/activate && python tools/skills/business-case/roi_engine.py --json '{
     "prospect": {"name": "<name>", "use_case": "<use_case>", "industry": "<industry>"},
     "inputs": { ... confirmed volume/cost inputs ... },
     "assumptions": { ... merged assumptions from benchmark YAML ... }
   }'
   

   Parse the JSON output containing results, formatted, interpretations, traces, and scenarios.

5. Run validate_model.py on the output. If any "fail" status findings, warn the user before proceeding.

6. Run sheets_builder.py to create .xlsx:

   source .venv/bin/activate && python tools/skills/business-case/sheets_builder.py --json '{
     "calc_output": { ... full roi_engine output with inputs_used and assumptions_used ... },
     "prospect_name": "<name>",
     "use_case": "<use_case>",
     "industry": "<industry>",
     "benchmarks": { ... parsed contents of the benchmark YAML (read the file first, then inline the dict) ... },
     "findings": [ ... validation findings ... ]
   }'
   

   calc_output is a required wrapper key. Do NOT pass the roi_engine JSON at the top level. The script does data["calc_output"] and raises KeyError: 'calc_output' if the key is missing. Always nest the full roi_engine output dict under "calc_output": { ... }.

7. Upload .xlsx to Drive via workspace-mcp create_drive_file:

   • file_name: "ROI Calculator - <Prospect>.xlsx"
   • mime_type: "application/vnd.google-apps.spreadsheet"
   • fileUrl: "file:///Users/nick.yebra/tmp/business-case/<prospect-slug>/roi-calculator.xlsx"
   • user_google_email: "[email protected]"

8. Get shareable link via get_drive_shareable_link.

Store ROI headline metrics and Google Sheet URL for Phase 3.

Phase 3: ASSEMBLE + CONFIRM

Step 1: Load Reference Files

Read these reference files for content generation:

• references/slide-content-formulas.md (per-slide data sources and content requirements)
• references/pricing-rate-card.md (pricing calculation formulas and SKU tables)
• tools/skills/create-presentation/references/scr-framework.md (SCR methodology for Slide 2)

Step 2: Synthesize Slide Content

Generate content for each of the 5 automated slides:

Slide 1 (Cover): Company name in title. Format: "TRANSFORM {COMPANY}'S CUSTOMER EXPERIENCE WITH AI AGENTS". Logo URL presented to user for manual insertion.

Slide 2 (Engagement Overview): Single slide with 3 paragraph subsections following compressed SCR framework:

• BACKGROUND (Situation): Customer's world in their language. Industry context, current operations, scale metrics. Source from company-researcher + edgar-researcher + obsidian-analyzer.
• COMPLICATION (Conflict): The problem and why now. Gap between strategy and execution. Source from transcript-analyzer (pain points) + salesforce-researcher (deal context). Use the gap formula: "[Priority X] is a stated top priority, yet [evidence Y] suggests the current execution isn't keeping pace."
• SOLUTION (Resolution): How Regal solves it. Initial use case, expected outcomes. Source from org-context (use-cases.md, differentiators.md). First mention of Regal here, not in BACKGROUND or COMPLICATION.

Text length constraints: 2-3 sentences per subsection, 120-180 words total across all three sections.

Slide 5 (Pricing Overview): Calculate estimated costs using intake inputs + rate card formulas. Present as labeled key-value pairs, one per line:

• Platform fee: $X/year
• AI Phone Agent: $X/min x Y monthly mins = $X/year
• Telephony ({SIP|VoIP}): $X/min x Y monthly call volume = $X/year
• Support tier: $X/month
• Estimated annual total: $X
• Discount tier ({commit level}): {X}%
• Discounted annual estimate: $X
• Minimum commitment: $X/quarter during IOT
• Initial term: {X} days

Slide 6 (ROI): Headline metrics from roi_engine.py output + Google Sheet URL. Show 3-5 headline metrics (e.g., "Annual savings: $X", "ROI: X%", "Payback period: X months"). Metrics vary by use case per references/slide-content-formulas.md.

Slide 11 (Close): Prospect-specific next steps. 3-4 numbered steps with timeframes, key contact name, action-oriented tone. Reference sales-process.md Stage 3 exit criteria for structure.

Step 3: Generate Order Form Checklist

Generate the order form input checklist from references/order-form-checklist-template.md with pre-filled values from research and intake (Section 1 values populated, Sections 2-3 as template).

Step 4: Present for Approval

Present ALL content for user review:

• All 5 slide contents with clear labels
• Pricing calculation breakdown showing formula inputs and results
• ROI headline metrics
• Logo URL (for manual insertion)
• Order form input checklist

Wait for explicit user approval. Accept revisions to any slide content, pricing adjustments, or scope changes before proceeding to Phase 4.

Phase 4: GENERATE + DELIVER

Step 1: Load Template Schema

Read references/proposal-template-schema.md for objectIds, page IDs, and population strategy.

Step 2: Copy Template

Copy the proposal template via workspace-mcp copy_drive_file:

• File ID: 1El0FEUrqw5pzjtZguVcWoFOruqEEnq2lZHfA3w4TFDU
• New title: Regal Proposal - {Company Name}
• user_google_email: "[email protected]"

NEVER edit the template in place. Always copy first.

Step 3: Build Batch Update

Build the batch_update_presentation request. All content must be plain text (strip markdown bold/italic markers, convert bullets to newlines, preserve numbers/currency).

Slide 2 text box resize (must precede replaceAllText): The BACKGROUND, COMPLICATION, and SOLUTION text boxes are header-sized (~0.4" tall). Adding paragraph content via replaceAllText will overflow. Before any replaceAllText calls, include updatePageElementTransform requests for the 3 SCR text boxes (g3cd20496ae9_0_901, g3cd20496ae9_0_904, g3cd20496ae9_0_905) to expand their height. Use RELATIVE transform mode with scaleY: 10.0 to safely multiply the current height without needing exact absolute coordinates. See EMU Dimensions section of references/proposal-template-schema.md for estimated dimensions and resize targets.

Target strings per slide (from references/proposal-template-schema.md), using replaceAllText with pageObjectIds scoping:

Slide 5 pricing table: The 13x4 table requires per-cell replaceAllText calls. Use the hardcoded cell values from the Known Cell Values section of references/proposal-template-schema.md as containsText targets. Do NOT attempt to read cell text at runtime; the get_presentation API does not expose table cell contents. Each cell's exact current text (Slavic 401k sample data) is documented in the schema. Replace: Col 2 (Unit Price) with prospect rates, Col 3 (Projected Volume) with intake volumes, Col 4 (Projected Spend) with calculated values. Also replace the 1x1 header table text 12-Month, $100k Commitment (Post POC) with prospect-specific terms.

Slide 11 special character: The template uses \u000b (vertical tab / soft return), not \n. The containsText must include \u000b to match.

Do NOT touch slides 3, 4, 7, 8, 9, 10, 12, 13, 14 (static or V2).

Step 4: Execute and Deliver

1. Execute batch_update_presentation with user_google_email: "[email protected]"
2. Get shareable link via get_drive_shareable_link
3. Present final deliverables:

Proposal deck: [Google Slides URL]
ROI calculator: [Google Sheet URL]

Logo URL for manual insertion into slide 1: [URL]

Order Form Input Checklist:
[checklist text]

Next steps:
- Insert the company logo into slide 1 (drag the image from the URL above)
- Review slides 8 (GANTT) and 10 (System Map) for V2 content
- Move an order form template and fill the Volume Worksheet tab

Error Handling

Individual Research Agent Failures (Phase 2)

• salesforce-researcher fails: Proceed without CRM data. Flag missing deal stage, ACV, and contact data. Pricing slide (5) relies on user intake inputs, not SF data, so it still works.
• obsidian-analyzer fails: Proceed without prior account context. Slide 2 (Engagement Overview) draws more heavily from company-researcher and transcript-analyzer outputs instead.
• transcript-locator or transcript-analyzer fails: Proceed without transcript data. Volume estimates rely on intake inputs. Flag that pain points and use case details may be less specific.
• company-researcher fails: Proceed without web research. Flag that slide 1 logo URL is unavailable and slide 2 Situation subsection may lack industry context. Slide 11 (Close) uses deal context from other agents.
• edgar-researcher fails or returns "Private company": Normal for private companies. Proceed without SEC data. No slides depend exclusively on edgar output.

Pipeline-Level Failures (Phase 4)

• Template copy failure: Present template ID 1El0FEUrqw5pzjtZguVcWoFOruqEEnq2lZHfA3w4TFDU for manual copy. Continue with remaining deliverables (ROI sheet, checklist).
• Batch update failure: Report which slides succeeded/failed. Provide the Drive file ID so user can manually edit remaining slides.
• ROI engine failure: Continue without slide 6 ROI content. Note the gap. Present rate card pricing without ROI metrics.
• workspace-mcp unavailable: Fall back to saving all content as an Obsidian document via obsidian_append_content at Regal/Proposals/{Company Name} - Proposal Content.md. Include all slide content, pricing, and checklist in structured markdown.
• Logo not found: Note in slide 1 content that no logo URL was found. User can source it manually.

Critical Rules

• All workspace-mcp calls require user_google_email: "[email protected]"
• NEVER edit the template in place; always copy first via copy_drive_file
• NEVER use Write tool to save to Obsidian (use obsidian_append_content)
• NEVER dispatch sub-agents with run_in_background: true (MCP tools unavailable in background)
• Dispatch all 5 research agents in a single message for parallel execution
• Read tools/references/output-persistence.md for the persistence pattern and fallback protocol
• ROI engine runs via direct Python call, NOT via nested /business-case skill dispatch
• All slide content must be plain text for replaceAllText payloads (strip markdown, bullets to newlines, preserve numbers/currency)
• Slide 2 SCR text: 2-3 sentences per subsection, 120-180 words total
• No mention of Regal in BACKGROUND or COMPLICATION subsections (SCR rule: first Regal mention in SOLUTION only)
• NEVER use obsidian_batch_get_file_contents for org-context reads (Step 2b): the 5 files total 100K+ chars combined and will overflow the batch token limit. Use 5 individual obsidian_get_file_contents calls instead.
• Wrap roi_engine output under calc_output key for sheets_builder.py: the script requires data["calc_output"] and raises KeyError if you pass the roi_engine output at the top level.