Overview

Snapshot for Adrienne · generated —

Pipeline by stage

Weighted · USD

This week

Calls · Emails · Meetings

Top deals to watch

Next-step dated

Deals

From HubSpot snapshot · owner = Ian Grady

Projects

Goals, initiatives, portfolio work · from Asana (pending PAT)

Activity

Recent calls, emails, meetings

Autoresearch

Karpathy-grade prompt iteration loop · Claude writes [AR_CONTEXT] notes, HubSpot Prospecting Agent drafts, ledger measures · —

Plan doc

Funnel

Notes → drafts → sent → replies → meetings

Infrastructure health

Smoke + dedup tests

Prompt-version leaderboard

Reply rate by version · winner stays, loser gets mutated

Runs

Per-day breakdown by research method

How autoresearch works here

Two skills, one HubSpot, one ledger

The pattern. Autoresearch is a Karpathy-style prompt iteration loop. Claude writes a research note onto a HubSpot contact, HubSpot's Prospecting Agent drafts an outbound email using that note, the ledger measures what happens (sent, reply, meeting), and the next run mutates one variable to try to do better. Winner stays, loser gets replaced.

Two autoresearch skills feed it.

Cloud autoresearch (in HubSpot): the Prospecting Agent + Breeze AI inside HubSpot's UI. Its Selling Context says "Use the latest [AR_CONTEXT] note as primary research source." When it drafts an email, it reads the most recent [AR_CONTEXT] note on the contact and grounds the email in that signal.
Claude Code autoresearch (local on Personal MBP): the skill at 00_AIOS/Skills/autoresearch.md Mode B. When triggered, it pulls real CRM context (engagement timeline, prequalification framework, stakeholder names) and writes a structured [AR_CONTEXT] note via Zapier MCP onto the target contact in HubSpot.

How they connect.

Claude Code writes the [AR_CONTEXT] note → Zapier MCP hubspot.create_engagement → contact's HubSpot timeline.
Ian (or a scheduled play) enrolls the contact in the HubSpot Prospecting Agent.
HubSpot's Prospecting Agent reads the latest [AR_CONTEXT] note, drafts an email grounded in the signal, queues it for review.
Ian approves in HubSpot "Ready for review" → email sends from ian@greenplaces.com.
Outcome (sent, reply, meeting) is logged into the autoresearch SQLite ledger in the vault.

Where the data lives.

Source of truth: SQLite ledger at ~/Obsidian/Ian-Vault-Lab/outputs/agent-handoff/autoresearch-ledger.sqlite (three tables: experiments, smoke_runs, dedup_tests).
Aggregate snapshot for this dashboard: regenerated manually via build-dashboard.py in outputs/agent-handoff/autoresearch-dashboard/, written to public/data/autoresearch.json, committed to git, deployed by Cloudflare.
What this tab reads: only the static aggregate JSON above. Per-experiment detail (contact emails, draft bodies) stays in the vault SQLite — never on a public deploy.

Currently live. Run 1 sandbox (3 self-sends, 2026-05-03) and Run 2 Trex (1 send fired, 2 stalled at note-only, 2026-05-04). Prompt variants v0.1.0 baseline, v0.2.0 tighten-length, v0.3.0 stakeholder-map are staged in outputs/agent-handoff/autoresearch-prompt-variants/; swap-in via Comet pending.

Admin

Feedback · integrations · setup · change log

Restricted

Enter admin password to continue.

AI feedback summary

gpt-4o-mini · top themes from open queue

Send feedback to Harper

Goes to the build thread · I'll ping you when shipped

Recent submissions

Latest 100 · auto-loaded

Integrations

Every tool wired into the Ledger or Harper

Setup & maintenance

How this is wired

Making changes to this dashboard

Three lanes: Claude Code, Cloudflare, HubSpot

The default workflow. Open Claude Code in ~/the-ledger. Tell it what you want. It opens a branch + PR. Glance at the diff. Click merge. Cloudflare auto-deploys within 1–2 minutes. If it's wrong, git revert and back to the previous version.

Claude Code handles (your terminal, in ~/the-ledger):

UI changes (HTML, JS, CSS in public/)
Backend functions (functions/api/*.js)
Database migrations (migrations/*.sql)
Tests (test/*.test.mjs)
Branch + commit + push + open PR via gh

Cloudflare handles (no manual touching of the dashboard required):

Hosting the static site at the-ledger-ak5.pages.dev
Running the backend functions (serverless edge)
D1 SQLite database (where structured data like feedback + future API tokens live)
Secrets storage (SESSION_SECRET, ADMIN_PASSWORD, OPENAI_API_KEY)
Auto-deploying every push to main within ~1 minute

HubSpot handles (the actual CRM data):

Source of truth for deals, contacts, companies, engagements
Connected to the Ledger via Zapier MCP using ian@greenplaces.com
Refresh path: Claude pulls HubSpot data via MCP → regenerates public/data/snapshot.json → commits → Cloudflare deploys

Common changes, where to start:

"Add a column to the Deals tab" → Claude Code, edits public/index.html + public/assets/app.js
"Refresh HubSpot data manually" → Claude Code, run npm run refresh-snapshot (lands when the refresh runner PR merges)
"Add a new feedback category" → Claude Code, edit the dropdown in public/index.html
"Change the admin password" → terminal in ~/the-ledger, run npx wrangler pages secret put ADMIN_PASSWORD --project-name=the-ledger
"See what auto-deployed today" → daily digest GitHub issue (lands when PR #6 merges)

What NOT to do (the things that wasted today):

Don't push directly to main. Branch protection blocks it; auto-deploy fires off main.
Don't bundle multiple unrelated changes into one PR. Smaller is faster to review and easier to revert.
Don't try to make HubSpot push to the Ledger directly. The flow is always Claude → MCP → vault → snapshot.json → deploy.

Change log

Every deploy auto-logged