Sales

Personal, local-first B2B sales research-and-outreach tool. Every factual claim in every draft traces to a verified evidence row. Drafts are critiqued against a user-owned principles file. Every revision is preserved.

Requirements

• macOS with the claude CLI installed and logged into a Claude Max 20 account
• Node.js 20.9+, pnpm

Setup

pnpm install
pnpm db:generate   # idempotent after first run
pnpm db:migrate
pnpm dev

Open http://localhost:3000.

Workflow

1. Create an account. Add the target company.
2. Gather evidence.
   • Paste URLs + raw text into the Evidence tab, OR
   • Click "Run auto-research" to have the Claude CLI research the account via WebFetch + WebSearch.
3. Audit evidence. Click "Run extraction audit on pending" — each fact is checked against its snippet. Disputed rows show a reason and a suggested correction.
4. Add a contact (optional; sets the buyer archetype for the drafter).
5. Create a sequence. Pick channels per touch (email / LinkedIn).
6. Draft each touch. The drafter uses only verified evidence. The validator ensures every cited claim is a verbatim substring of its evidence snippet. If the draft fails validation, it retries once with a correction message.
7. Run critics. Three critics score each draft:
   • Skeptical Buyer (Sonnet) — "Would I delete this in 2 seconds?"
   • Sales Coach (Sonnet) — scores against every principle in data/principles.md.
   • Writing Editor (Haiku) — concision, AI-tell phrases, active voice.
8. Accept rewrites. Each accepted rewrite creates a new immutable revision; prior revisions and their critiques are preserved.
9. Export. Download .eml files for email touches and .txt for LinkedIn. Touch 1 is copied to clipboard.

Key files

• data/principles.md — the Sales Coach critic rubric. Edit as your tactical bar evolves.
• data/icp.md — ICP brief. Read by the drafter on every touch. Fill in before your first draft.
• skills/ — Claude Code-compatible skill files used by the CLI subprocess.
• docs/superpowers/specs/ — full design spec (v2).
• docs/superpowers/plans/ — implementation plan.

Architecture

Next.js 16 App Router + SQLite (Drizzle ORM) + claude CLI subprocess for every LLM call. No Anthropic API key is used — the CLI authenticates via the owner's existing Max 20 OAuth session.

Three layers with strict contracts:

1. Evidence (spine): typed, append-only store. Every fact has a source_url, snippet (≤1500 char verbatim excerpt), and extractionStatus (pending_audit | verified | disputed). Only verified rows can be cited in a draft.
2. Drafting: pulls verified evidence, calls Claude, emits cited_evidence_ids + supporting_spans (verbatim substrings of snippets). Validator rejects any span that isn't a substring; one-retry loop on failure.
3. Critique: 3 parallel critics produce structured findings. Accept/reject per suggestion. Accepting a rewrite creates a new immutable touch_revisions row; prior revisions remain intact.

Scope

v1 MVP. Out of scope:

• SMTP/Gmail sending (copy-paste the .eml into Gmail).
• CRM sync.
• Multi-user / SaaS.
• Deep Research paste parser, Perplexity MCP, GPT-5 critic — deferred to v1.1+.

See docs/superpowers/specs/2026-04-17-sales-tool-design.md §10, §11.

Tests

pnpm test
pnpm typecheck
pnpm build