sales-call-prep

Run the ten-stage sales call-prep route on a company and produce a formatted PDF brief (account brief, earnings-call pain, contact role, discovery questions, opener, buying committee, objections, competitors, battle card, follow-up plan). Use when the user gives a company name for call prep, account research, or a battle card.

Call-Prep Route

The user gives you a company. You run all ten stages in order and return a PDF. They should never have to write a research prompt.

Support files (load on demand):

• references/prompts.md - the ten prompts, stop-rule fallbacks, and "done" criteria. Read first.
• references/seller-profile.md - the user's product, proof points, pricing, method.
• scripts/build_brief_pdf.py - JSON payload to PDF. Schema is in its header docstring.
• assets/sample_payload.json - complete fictional payload showing the exact shape.

When to Use

• "Prep me for a call with X", "research X for a sales call", "brief me on X"
• "Build a battle card for X", "run the route on X"
• A bare company name in a sales context

Do not use for general company research with no sales purpose.

Procedure

1. Intake (ask at most once). Required: company name. Optional: contact name and title, deal type. Read references/seller-profile.md; if STATUS: TEMPLATE, ask ONE batched question for the fields marked (*) and offer to save the answers into that file. If the user says just run it, proceed and tag solution-fit content "inferred". Resolve the company with a search, verify the domain from results (never guess it), state the identity you chose in one line, and ask only if the name is truly ambiguous.
2. Load references/prompts.md and run stages 01 to 10 in order. Carry findings forward: discovery questions trace to stages 01-03, the opener to a dated finding, the battle card to stages 06-08, the follow-up to everything. Budget about 12-20 searches or page extractions; search each distinct item separately; prefer primary sources (company site, IR pages, SEC filings, press releases); extract full pages rather than relying on snippets.
3. Apply the stop rule at every stage. If a source is gated or missing, record it in gaps[] with a concrete ask ("Paste her last 3 posts"). Never bypass logins or paywalls. Never invent names, quotes, prices, metrics, or customer results; write "Not public" or "Needs rep input". Tag each claim sourced (with a source id) or inferred.
4. Quotes. Stage 02 needs exact wording: short verbatim excerpts (at most two sentences) with speaker, call, and link. If exact wording is unavailable, drop the quote and tag "inferred".
5. Write the payload to ~/call-prep/<company-slug>/payload.json, matching assets/sample_payload.json. Set meta.prepared_on to today and meta.sample to false. Write snapshot.takeaways last.
6. Validate: python3 scripts/build_brief_pdf.py <payload.json> --validate (run from the skill directory, ~/.hermes/skills/sales-call-prep/). Fix real warnings.
7. Build: python3 scripts/build_brief_pdf.py <payload.json> ~/call-prep/<company-slug>/<company-slug>-call-prep-<YYYY-MM-DD>.pdf. If reportlab is missing, pip install reportlab.
8. Deliver. Attach the PDF if the current platform supports file attachments; otherwise print the full path. Reply briefly: the three findings most likely to change the call, then the gaps that need the user's input. State that the follow-up email is a draft and nothing was sent.

Pitfalls

• Wrong-company research from a guessed domain. Verify from search results.
• Turning a fact into a bigger claim ("hired a VP of Data" is not "building an AI team"). Label inference.
• Inventing proof points or competitor pricing. Only use the seller profile or cited sources.
• Long verbatim passages from transcripts. Keep excerpts short.
• Payload item counts off: 3 pain points, 5 discovery questions, 3 objections, 5 battle-card entries.
• Unicode symbols outside Latin-1 in text; the script replaces them, so avoid them.

Verification

• --validate prints no unresolved source ids or count warnings.
• Every empty stage has a matching gaps[] entry.
• Open one or two rendered pages (pdftoppm -r 60 -png file.pdf out) and check for clipped text or orphaned headings.
• The PDF page 1 shows the snapshot, meta table, and route status; the last page lists gaps and sources.