Ops Manager Day One — access, machine setup, and the run-of-show

1 Read this first

Today has four parts, in this order: you grant access (30 minutes, before he arrives), he sets up his machine on a screen share (about an hour), you give him context (30 minutes), then you do one real example in each of the six areas together, each with a definition of done. Send him the Start Here page this morning. It has the six links in plain words and the definition of done. He reads the cheat sheet tonight, not this morning.

The one fact that shapes access. The workspace repo carries the database connection file inside it. Whoever has a copy of the repo can reach the brain, and the brain holds the credential vault. So the copy of the workspace is the master key. Today that copy is the setup bundle (section 4), not GitHub: the GitHub copy is two months stale (last push 13 Jul) and cannot be pushed until three files over 100 MB are removed from history. Offboarding means deleting his copy and rotating the database password the same hour.

What he does not get today: anything sales-side (Kixie, Instantly, Stripe, Calendly), the scheduled jobs that run on your Mac, and the iMessage bridge. Those stay on your machine.

2 Before he arrives: what to grant (about 30 minutes)

3 Credentials: how to hand them over

There are two kinds. Keep them separate in your head.

Human-facing secrets (typed into a browser): the rooms passcode, the B2C board key, the Sales Center key, the DripJobs logins. Put these in a shared password-manager vault (1Password or Bitwarden, a vault named "St. Pierre Ops") and share the vault with him. If that is not set up by 9:30, read them aloud on the call and he types them once; each page remembers its key. Never send them in Slack, email or a text.

Machine-facing secrets (used by scripts): every API token lives in the credential vault inside the brain, table credentials_vault, read with brain.get_credential(). He never copies a token into a file. Once he has database access, his Claude reaches every token the same way yours does.

Database access: pick one

Pick A today, C inside the first two weeks. Reversible, and the alternative is a blind operator. Write the rotation date in the ledger when you grant it.

4 His machine: from a blank Mac to running the systems (about an hour, on screen share)

Do these in order. Every step ends with a check. Stop at a failed check; do not skip ahead.

1. Command-line tools and Homebrew

xcode-select --install
/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"

Check: brew --version prints a version.

2. The tools the scripts expect

brew install git gh python@3.12 node ffmpeg whisper-cpp

Check: python3.12 --version, node --version, ffmpeg -version | head -1, whisper-cli --help | head -1 all print.

3. Python packages (always python3.12, never plain python3)

python3.12 -m pip install --break-system-packages psycopg2-binary requests playwright pyyaml
python3.12 -m playwright install chromium

Check: python3.12 -c "import psycopg2, requests, playwright; print('ok')" prints ok.

4. Claude Code and the desktop app

npm install -g @anthropic-ai/claude-code
brew install --cask claude

Then run claude once and sign in on his account. Check: claude doctor reports healthy, claude --version prints 2.1 or newer.

5. The workspace, from the setup bundle (not from GitHub)

You build the bundle once on your Mac and hand it over by AirDrop, an external drive, or Google Drive shared to his company account only (never a link share). Both of you delete it after the install:

zsh /Users/akash/Workspace/Claude/.codex/references/ops-manager-setup-bundle.sh

It writes ~/Desktop/stpierre-ops-manager-bundle-<date>/ with the whole workspace as a git bundle (history included, about 15 GB), the skills, the memory notes, the deploy-gate shim, and an install.sh. On his Mac, from inside that folder:

zsh install.sh

That one script does steps 1 to 9 of this section, in order, and stops at the first failed check. If you prefer to walk it by hand, the manual steps continue below. Check: ls ~/Workspace/Claude/AGENTS.md exists.

Why not git clone from GitHub: the GitHub copy stopped receiving pushes on 13 Jul (39 commits behind), and three tracked files over 100 MB (a 1 GB zip and two others under _scratch and market research) block any push until they are removed from history. That cleanup is a separate task; it does not block today.

6. The path fix (do not skip)

sudo ln -s "$HOME" /Users/akash
ln -s ~/Workspace/Claude ~/Documents/Claude

Check: ls /Users/akash/Workspace/Claude/.codex/references/brain.py exists.

7. Prove the brain works

python3.12 - <<'EOF'
import sys; sys.path.insert(0, '/Users/akash/Workspace/Claude/.codex/references'); import brain
print(brain.query_one("select now() as ts"))
print(brain.cols('clients')[:5])
c = brain.get_credential('stpierre-b2c-board', scope='stpierre', accessed_by='<his name>', context='day one setup check')
print('vault ok:', bool(c and c.get('secret_value')))
EOF

Check: a timestamp, five column names, and vault ok: True. If the vault line is False, his database access model is wrong (section 2).

8. The deploy gate shim (already installed by install.sh; manual path below)

Files: ~/.local/sp-bin/npx and ~/.local/bin/sp-deploy-gate-check. He puts them in the same paths, then:

chmod +x ~/.local/sp-bin/npx ~/.local/bin/sp-deploy-gate-check
echo 'export PATH="$HOME/.local/sp-bin:$PATH"' >> ~/.zshrc && source ~/.zshrc

Check: which npx prints the sp-bin path. Without this, a deploy from his Mac skips the gate.

9. Skills and memory (already installed by install.sh; manual path below)

• ~/.claude/skills/ (26 skills). He copies the folder to the same path. These are the how-to scripts his Claude follows: "build the FR pack", "stage crawl space for tos", "run the launch QA".
• ~/.claude/projects/-Users-akash-Workspace-Claude/memory/ (275 one-fact notes: your rulings and the lessons). Review it for personal notes first, then he copies it to the same path on his Mac. Without it his Claude starts every session cold. The brain's session handoffs are the shared memory either way.

Check: he opens claude inside ~/Workspace/Claude and asks "what are the six areas the ops manager owns?" The answer should cite the cheat sheet.

10. Connectors in his own Claude desktop app (only as needed, never today): Google Workspace, Notion, Zapier, Slack, Descript. These are per-account. The Meta ads connector is optional; the scripts reach Meta through the vault.

Never on his Mac: the com.stpierre.* scheduled jobs (47 of them run on yours; a second copy would double every send), the MyCRMSIM iMessage bridge, anything under Scheduled/. Voicebox, Descript and Higgsfield wait for the first content build.

5 What was already built for this in July, and what to reuse

• stpierre-builder repo plus Codespace (30 Jun to 14 Jul): a cloud box with 12 secrets injected at boot and a database role locked out of the vault. Built for Collins as a builder, never handed to anyone. Reuse the idea (secrets as environment, never in files) and the smoke-test habit (verify.sh must show a non-zero count, not just "ran"). Do not reuse the box for the ops manager: it holds none of the workspace.
• The builder onboarding launcher (10 Aug, 02-Areas/Team/builder-onboarding-launcher-2026-08-10.html) and the two lessons it found: a scoped role with row security and no policies sees zero rows and still reports PASS, and a GHL seat is the one login worth giving even under an API-first rule. Both are baked into this doc.
• Builder Delivery Handbook and Delivery Operating Manual (3 Sep, team.stpierre.ai/onboarding/handbook and /operating-manual): the roles split and the daily sweep. Still current; the cheat sheet points to them.
• The .devcontainer in the workspace (setup.sh, verify.sh): the Codespace recipe. The install.sh in the bundle is its Mac equivalent.

6 Context: the 30-minute talk-track

• 1"You own six things. I sell." The six, one line each, from the top of the cheat sheet.
• 2"Each area has one surface you live in." Board, ledger, client lane, scoreboard, bugs board. Everything else is reference.
• 3"Your day has three clock points." By 9:15 prove yesterday's leads landed. By 3 pm hand me one message per client. At shutdown log every clunk.
• 4"Three rules that are not up for debate." Delivered means the client's board, not our label. Built does not mean running. Anything seen three times becomes a system, not a fourth fix.
• 5"Five things stay with me." Money, messages sent as me, public-facing changes, account access, deleting data. Everything else: do it, ledger it, report it.
• 6"Your first week is not a rebuild." Own the bugs board, get the client messages moving, close the three conversation loops.

7 The schedule

8 Example 1 · Pages (30 minutes)

Goal. Run the readiness sweep on the whole page fleet, find the three known defects, log them.

• 1Open the cheat sheet, Area 1, and read the fleet table together (two minutes).
• 2In Terminal, run the readiness sweep and read the output aloud: page name, Lead Ads terms, token visibility, for each page.

python3.12 /Users/akash/Workspace/Claude/01-Projects/StPierre-GTM/fr-transition-2026-08-26/state-fleet/check_page_readiness.py

• 1Find the three defects: Indiana still on the old name; Ohio with a client row missing; the privacy page every lead form points at returning an error. Confirm the last one:

curl -sI https://legal.stpierre.ai/privacy-texting | head -1

• 1Log each as an issue, owned by him. In Claude: "Add three rows to the issues table: source ops-manager, severity critical, status open, owner <his name>, one per defect, with the page id in the detail." Claude writes them through the brain and reads them back.
• 2Save the sweep output to 02-Areas/Team/ops-manager-log/2026-09-07-pages.txt.

Definition of done. Three issue rows with his name. The sweep output on disk. You decide at 3:45 whether Indiana gets re-submitted and whether Ohio gets stood up now.

What only you can do here. Create, claim, rename, accept terms, connect in GHL. He prepares, you click. Say the GHL connection reminder out loud in every launch ask.

9 Example 2 · Content (30 minutes)

Goal. Read what won, and write the brief for the next foundation repair pack.

• 1Open the Creatives tab. Set All time. Sort by cost per qualified lead. For the top three: slug, spend, qualified, cost per qualified, the hook words, the surface on screen.
• 2Open Drive, "Proven B2C Ads by Niche", foundation repair. Match the top three to their master files.
• 3Open 01-Projects/StPierre-GTM/ready-to-deploy/2026-08-13--fr-voiced-pack/contact-sheet.html. This is what a finished pack looks like. Read the README's rulings.
• 4Write the brief: three hooks with their numbers, which master to re-voice first, which market it is for. Save as 01-Projects/StPierre-GTM/ready-to-deploy/2026-09-08--fr-pack-brief.md.
• 5In Claude, say "build the FR pack" and watch the skill trigger. Stop at the plan. Voicebox is not on his Mac yet; the build happens on the first content day.

Definition of done. The brief saved with three hooks and their numbers. You name the relight market.

Standards to say out loud. Never fewer creatives. One campaign per market, budget at the campaign level. Per-niche ad accounts, checked before every launch. Light only. No adjectives.

10 Example 3 · Conversations (30 minutes)

Goal. Work five simulator patterns, then diagnose one stuck conversation in the right order.

• 1Open the conversation simulator. Homeowners tab. For five cards, type what he would say, or "nothing". Copy all answers, paste them to you. You react live; that is the script forming.
• 2Open the B2C Board. Pick one card in "Needs reply" or "Unreachable" with a phone.
• 3Diagnose in this order, no theories first. (a) The send guard's log for that lead, which says whether each automated message was allowed and why:

python3.12 - <<'EOF'
import sys; sys.path.insert(0, '/Users/akash/Workspace/Claude/.codex/references'); import brain
lead = '<lead id from the card>'
for r in brain.query("select decided_at, kind, channel, allowed, reason from sp_send_guard_log where lead_id = %s order by decided_at desc limit 10", (lead,)): print(r)
EOF

(b) GoHighLevel → the contact → the workflow's execution log. A step marked "Executed" means the system tried, not that anyone received it. (c) The live contact record. (d) Only now, one sentence on the cause.

• 1Write the cause and the next action as a note on the card.

Definition of done. Five answers in your hands. One card with a cause and a dated next action.

Rules to say out loud. One script for everyone. Never the same link twice in 48 hours. Testing on your own number fakes a loop; the fixture contact is id 88. Nothing is armed on a blanket go.

11 Example 4 · Lead transfers (30 minutes)

Goal. Prove where every lead from yesterday actually is.

• 1Open the lead ledger. Filter to yesterday. List every lead with a name and a phone.
• 2For each, the delivery rows (one per place the client receives leads):

python3.12 - <<'EOF'
import sys; sys.path.insert(0, '/Users/akash/Workspace/Claude/.codex/references'); import brain
print(brain.cols('sp_client_lead_transfers'))
for r in brain.query("select lead_id, channel, status, last_error, sent_at from sp_client_lead_transfers where created_at > current_date - 1 order by created_at desc"): print(r)
EOF

• 1For each Level King or AlphaLift lead, prove the card: DripJobs global search by name on the client's board (your Chrome, screen shared). A card exists or it does not.
• 2Sort them: proven (card seen), delivered-unproven (sent row, no card seen), stuck (error or no row). Donny Spurgin is the live example of stuck: a hold since 3 Sep, watchdog paging daily.
• 3Stuck ones get an issue row or a nudge.

Definition of done. A three-column list for yesterday, saved to the ops log. Every stuck lead has an owner.

Rule to say out loud. Delivered is a timestamp. A text to your own phone is never a client delivery. The client's board is the proof.

12 Example 5 · Post-transfer (45 minutes)

Goal. Turn the drafts nobody sends into one message per client, in your hands by 3:00.

• 1Open the client lane. Read the red cards per client. Open each client's coaching draft (copy button).
• 2Check the board before asking: Austin's board for M Saiful Razzaque (waiting on client since 1 Sep), Dan's calendar for his reds.
• 3Draft one message per client. Portal-first line always ("mark it in your portal, or reply with one word"). Never a message as Akash.
• 4Hand both to you by 3:00. You send between 3 and 4. He logs the send.

Definition of done. Two client messages sent by 4:00, logged. The client lane reds have a dated next action.

The ruling you make today. May he send these directly from tomorrow, or does "he drafts, you send" stay?

13 Example 6 · Issue tracker (30 minutes)

Goal. Own the bugs board, log a clunk, ledger a fix.

• 1Open the bugs board. Filter to unowned. Pick ten cards that belong to areas he touched today.
• 2Set the owner on their register rows. In Claude: "Set owner to <his name> on the issues rows for these ten bug keys." Claude updates the rows and reads them back.
• 3Rebuild the board and refresh the hosted copy:

python3.12 /Users/akash/Workspace/Claude/01-Projects/CommandCenter/scorecard/pipeline_bugs_tracker.py

Then in Claude: "copy today's bugs tracker to the rooms worker as bugs-board.html and deploy through the gate."

• 1Log one clunk from today:

python3.12 /Users/akash/Workspace/Claude/.codex/references/clunk.py add <pattern-key> "<what broke, one line>"

• 1Ledger one thing he changed today with a rollback. In Claude: "Log a change: project ops-onboarding, what I changed, why, and how to undo it."

Definition of done. Ten cards with his name. One row in the clunk log. One ledger row with a reverse_how.

Rule to say out loud. If you cannot write the rollback, the fix is not designed. Three times in seven days is a build, not a fourth patch.

14 Skills and rails index (what already exists, by area)

Skills (his Claude follows these when he says the trigger phrase)

Database jobs (run inside the brain, on a schedule, on nobody's Mac)

Mac jobs (47, on your Mac only; he reads their logs in 01-Projects/CommandCenter/watchdog/logs/)

15 Definition of done

Day one is done when

• His Mac passes the install check (vault ok: True) and his Claude, opened in the workspace, answers "what are the six areas the ops manager owns?" from the cheat sheet.
• Every row in the grant table is sent and he has opened all six links on the Start Here page.
• One real artifact per area carries his name: three page defects logged; the next-pack brief saved; five simulator answers in your hands and one stuck conversation explained on its card; yesterday's leads sorted into proven, unproven, stuck; two client messages sent by you and logged; ten bugs owned, one clunk logged, one ledger row with a rollback.
• You have answered the eight rulings, each as a ledger row or a reconciler check.
• He says the six areas and the three clock points back to you without looking.

Week one is done when (Friday 11 Sep)

• He ran the 9:15 sweep alone five days running, with a saved log each day.
• Every lead this week with a name and a verified phone is proven on a client's board or has an owned issue row. Zero "delivered to nobody."
• Client messages went out every day by 4:00. No coaching draft sat unsent past 24 hours. Waiting-on-client alerts under 48 hours old only.
• The bugs board has zero unowned cards, every three-strike bug has a build task, and the board rebuilds nightly.
• Indiana, Ohio and the privacy page are fixed or ruled. The page readiness sweep runs clean.
• One foundation repair pack is staged paused with a contact sheet, and the organic rail decision is made.
• His simulator answers are mapped into the template table, the message mirror is fresh, and the nightly fixer is loaded.
• At least five clunks logged and at least one became a permanent fix.
• You have picked the date for the scoped database role.

16 Recommendations on everything else (options, the pick, the why)

17 The eight rulings you make today

• 1Database access: option A today, option C by when?
• 2Indiana rename: re-submit now?
• 3Ohio page: stand it up now, or wait for a buyer?
• 4Who creates the eight remaining state pages, and when?
• 5May he send client messages directly, or does "he drafts, you send" stay?
• 6May he close issue rows alone, and must a close carry a ledger id?
• 7Deploy the staged scripted Messenger sender, or keep the AI bots on autopilot?
• 8Which market gets the first foundation repair pack, and on what relight date?

Write each answer as a ledger row or a reconciler check the same day. A ruling that lives only in chat is the first recurring problem in the tracker.