BT Tree Care Copilot

What it is

A Cloudflare Worker that puts a private chat UI in front of Anthropic's Claude API, with the entire Tree Care campaign baked in: brand voice, ICP, all 24 LinkedIn posts (hooks + full copy), DM templates, Sperry case study facts, pricing, and the rules for what works (and what doesn't) with tree care owners.

It does what the ChatGPT Project was doing — except this one runs on infrastructure we control. No per-seat caps. No regional throttling. No "we've temporarily limited access to your conversations" errors. We own the IP, the UI, the data, and the model selection.

The numbers

What's in it

Build timeline (May 12, 2026)

Companion: the DM Check bookmarklet

Drag-to-bookmark-bar button that grabs your draft text from any LinkedIn DM composer (or any selected text on the page), opens the BT Tree Care Copilot's DM Check pre-filled, and gives you a strict V3 brand-voice review without leaving LinkedIn.

Get the bookmarklet

Use case: Lazar composing a reply in LinkedIn → highlights his draft → clicks 📩 BT DM Check in his bookmark bar → instant strict feedback + suggested rewrite. Cuts the "remember to check it" friction to zero.

The voice rules — V3 (current)

Warmth vs. flattery — the line

The most-broken voice rule on this campaign. The distinction:

• Warmth is in scope. Peer-to-peer friendly tone — the way one contractor talks to another. Examples: "Good seeing tree care work get a real spotlight." "Hope the spring rush isn't crushing you." "Appreciate operators who'll talk shop straight." "Wanted to connect."
• Flattery is out of scope. Transactional positive judgments about the prospect's work or business. Examples: "You've built a solid reputation." "Impressive operation." "Your work is solid." "Looks like a company that's been built the hard way."

The test the AI runs on every draft: "Does the line judge their work or business positively? If yes → flattery → cut. Does it acknowledge something without rating it? If yes → warmth → keep."

3-part opener structure

1. Observation — one specific, concrete thing from their profile/post/company. Neutral statement. Not a positive judgment.
2. Reason for reaching out — one short, transparent line. Why are we in their DMs? Frames us as value-for-them, not asking-for-them. Removes the "what does this stranger want" friction. Research benchmark: senior targets respond best when the message under-promises and reads as honest.
3. Low-friction question — open-ended, easy to answer. Not "do you want to chat" — something they actually have an opinion on.

Hard limits: max 2 sentences / ~60 words, zero pitch in first-touch, zero flattery, mandatory specific observation.

Model openers

Pattern A — observed post + reason + question:
"Saw your post about [topic]. Working with a tree care owner in Eugene on
the business side of the operation — wanted to connect with folks
thinking about [that same problem]. How's that showing up in your work?"

Pattern B — observed service + reason + question:
"Saw [Company] is doing [specific service] in [location]. Working with
another tree care owner in OR on the business side. How's [that side of
the work] been trending for you this season?"

Pattern C — credential + reason + question:
"Saw you're [credential, e.g., 'a Certified Arborist in the PNW'].
Working with tree care owners on the systems side of the business.
Curious how you're handling [seasonal swings / database re-engagement]?"

Daily 6pm PT audit — Cowork scheduled task

A Cowork scheduled task runs every evening at 6pm PT and does the following:

1. Fetch current system prompt via the admin endpoint (so the audit grades against whatever rules are in force at the moment of the run)
2. Scrape Lazar's LinkedIn activity from the last 24 hours via Chrome MCP (Jason's logged-in BOSSTORQUE Chrome session): sent DMs, comments on posts, posts published
3. Grade each artifact using the /brand-voice:brand-voice-enforcement skill — classify as BROKEN / WEAK / GOOD / EXCELLENT
4. Suggest rewrites for anything sub-GOOD using /marketing:draft-content and /sales:draft-outreach
5. Identify recurring failure patterns across the day and any new rule or banned phrase that should be added
6. Post the report to #tree-care-campaign via bt-notify with quoted examples, recommended rewrites, and the pattern of the day
7. Auto-update the system prompt via the admin endpoint if a new banned phrase or rule emerged
8. Redeploy this build doc with a fresh timeline entry so the team always has current state

Task ID: bt-tree-care-lazar-eod-audit. Notify-on-completion is on — Jason gets pinged in Cowork after every run.

Architecture

One file, ~3000 lines of code, ~32 KB gzipped. No build step. SheetJS loaded from CDN, otherwise no dependencies.

D1 schema

conversations (id, user_id, title, model, created_at, updated_at)
messages      (id, conversation_id, role, content, sent_by,
               input_tokens, output_tokens, cache_creation_tokens,
               cache_read_tokens, cost_usd, model, created_at)
feedback      (id, message_id, conversation_id, user_id, rating, comment,
               created_at, ai_classification, ai_recommendation, reviewed_at)
config        (key, value, updated_by, updated_at)
source_files  (id, filename, content, size_bytes, uploaded_by, created_at)

API endpoints

The Lazar workflow change

Today's live demo proved the value. Lazar's daily list of 30 LinkedIn URLs was un-scoreable without enrichment — bare slugs only. Using Chrome MCP to scrape each profile, we got:

• 17 strong fits (8–10) — owners of tree care companies in target geo
• 4 maybes (5–7) — borderline (combo landscape, consulting variants)
• 9 no fits (1–4) — clustered in 3 clean patterns:
  • 4 wrong title (Arborist / Ops, not owner) — fix: add Job Title filter in Sales Nav
  • 2 consulting arborists (off-ICP per Jason) — fix: skip anything with "Consulting" in title or company name
  • 3 wrong vertical (Catholic Church, Tribal Council, Commercial Landscape) — fix: verify tree/arborist in headline

Full graded report (deployed as a CF worker, cardless internal reference).

Status checklist

• Worker deployed at bt-tree-care-copilot.jason-8ce.workers.dev
• D1 database created and migrated (5 tables)
• Secrets set: ANTHROPIC_API_KEY · GATEWAY_PASSWORD · BT_NOTIFY_TOKEN
• Daily cron schedule active (0 15 UTC = 8am PT)
• Anthropic auto-reload configured ($10 trigger / $100 refill)
• bt-notify channel tree-care-campaign registered with BOSSTORQUE Ops webhook
• 3-pass QA completed — functional, edge cases, production readiness all clean
• Admin role enforced (Jason + Armen) — Lazar gets 403 on admin endpoints
• Shared workspace: all users see all team chats
• Feedback persistence + AI review + daily Slack report all working
• Help chatbot deployed (lower-right ? button)
• Lazar + Armen notified in #tree-care-campaign — migration on
• System prompt V3 deployed — warmth/flattery distinction, 3-part opener structure, reason-for-reaching-out rule
• Cowork scheduled task bt-tree-care-lazar-eod-audit active — daily 6pm PT audit
• ChatGPT workspace cleanup: Jason to remove Armen + Lazar seats today
• Run-now the scheduled task once to pre-approve Chrome MCP + bt-notify permissions so future 6pm runs don't pause

Productized IP angle

This pattern is reusable. Same codebase, different system prompt and knowledge base, ships a custom copilot for any client campaign:

• Sperry Tree Care (this build) — tree care vertical, OR/WA/ID/NV/AZ/CA/BC, 24-post organic campaign
• Next client — clone, replace system prompt + source files in the Admin UI, deploy as <client>-copilot.jason-8ce.workers.dev, ~30 minutes

Add-on revenue model: bundle "AI Campaign Operations Portal" into client retainers at $500–$2,000/mo. Five clients = $30K–$120K/yr of recurring revenue from an asset that gets built once. The full codebase is sellable IP at exit. The Admin UI in particular makes this productizable — client side can self-manage their voice + knowledge over time without engineering involvement.

Deploy / maintenance

Update the worker

cd /Users/Jason/Library/.../outputs/gateway
wrangler deploy

Trigger daily report manually (for testing)

curl -X POST -H "X-Auth: treecare-2026-bt" \
  https://bt-tree-care-copilot.jason-8ce.workers.dev/api/cron-test

Send a direct message via bt-notify

TOKEN=$(cat '/Users/Jason/My Drive (jason@bosstorque.ai)/4_Strategy & IP/Entrepreneurial Ideas and Assets/GiftCue/.bt_notify_token')
curl -X POST https://bt-notify.jason-8ce.workers.dev/send \
  -H "Authorization: Bearer $TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"channel":"tree-care-campaign","text":"...","source":"manual"}'

Query D1 usage

wrangler d1 execute bt-tree-care-copilot --remote \
  --command "SELECT user_id, COUNT(*) AS msgs, ROUND(SUM(cost_usd),4) AS cost
             FROM messages m JOIN conversations c ON c.id=m.conversation_id
             WHERE m.role='assistant' GROUP BY user_id"

Check recent feedback

curl -H "X-Auth: treecare-2026-bt" \
  https://bt-tree-care-copilot.jason-8ce.workers.dev/api/feedback/recent?days=7