Skip to content

Add torpedo skill — account penetration via Amplemarket MCP#11

Draft
joaompena wants to merge 5 commits intomasterfrom
add-torpedo-skill
Draft

Add torpedo skill — account penetration via Amplemarket MCP#11
joaompena wants to merge 5 commits intomasterfrom
add-torpedo-skill

Conversation

@joaompena
Copy link
Copy Markdown
Collaborator

Summary

Adds the torpedo skill (v1) — an agent skill that helps sales reps break into target accounts by combining deep research (Amplemarket MCP + web search) with strategic planning and personalized sequence creation.

This is a draft PR — we're still iterating on the skill with users and will refine based on testing and feedback.

See the Initial MVP thesis for full context on what we're doing here and why.

What the skill does

  • Phase 0: Builds context about the user's company, product, competitors, and target personas
  • Phase 1: Deep research on the target account — existing state, company intelligence, signals & timing, org chart and people mapping
  • Phase 2: Synthesizes research into actionable strategy recommendations with specific angles and contacts
  • Phase 3: Creates personalized multi-channel sequences in Amplemarket for each approved contact

Files

  • skills/torpedo/SKILL.md — the skill (114 lines, lean and intent-driven)
  • skills/torpedo/preferences.md — empty preferences file, populated per user over time as they use the skill

🤖 Generated with Claude Code

joaompena and others added 5 commits April 16, 2026 12:37
…lized sequences

New skill that helps reps break into target accounts by combining Amplemarket MCP tools
with web research to build strategic engagement plans and create bespoke outreach sequences.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Don't reference internal skill mechanics (phase names, skill name) in conversation
- Mark Phase 1 research dimensions as mandatory to prevent skipping context-gathering
- Search org charts by department over specific titles, use larger page sizes
- Keep Phase 2 recommendations concise and scannable, not walls of text
- Store preferences as flat dated bullets, not structured sections
- Add two notes: stay grounded (don't fabricate), write outreach copy naturally

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The skill is primarily used in Claude Cowork and chat where the skill
filesystem is read-only — preference writes don't persist between sessions,
so the preferences mechanism was misleading. Drop it entirely; the skill
now asks for user context at the start of each conversation and holds it
in working memory for the session.

Earlier commits on this branch preserve the with-preferences version if
we need to recover it.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The previous default was too complex for the model to follow reliably
(branching on connection accepted vs not, three follow-ups all 3 days
apart). New default is a flat 4-step sequence with varied spacing.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Expand the copy-writing note with more specifics drawn from real test
  runs: stop with em dashes, no humblebrags, no trying-too-hard cleverness,
  no templated openers/CTAs
- Add hedging guidance: frame inferences about prospects as tentative
  hypotheses, not hard facts. Empathy prompt: how will the recipient feel?
- Fix "knowledge basis" → "knowledge bases"
- Fix confusing "your LinkedIn profile about section" → clarify it's the
  prospect's About section being searched by the keywords filter

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant