# EXTERNAL-AGENT HANDOFF — paste into ChatGPT or phone Claude

> **What this is:** a self-contained context dump Sam can paste into a fresh ChatGPT / phone Claude session so the external agent can pick up where his Claude Code workspace left off.
>
> **Generated:** Tue Apr 28 2026 · 5:39 PM EDT (NY) · session 18 close
> **Source trail:** `outputs/2026-04-28_17-39_handoff_external-agent-pickup.md` · workspace `C:\Users\ztrei\OneDrive\2. Hook Street\05. 2026 BH` · workspace commit `a4a5d00` · CM_Invoice_System commit `d979100`
> **What to do with the response you get:** copy the external agent's full reply, paste back into Claude Code with the line *"Here's what [ChatGPT / phone Claude] said — what should I do with it?"*

---

## 📋 PROMPT BODY — copy everything below this line

---

I'm Sam Treitel — owner of Hook Street Capital LLC. I work primarily out of a Windows + OneDrive workspace at `C:\Users\ztrei\OneDrive\2. Hook Street\05. 2026 BH\` with Claude Code (the CLI, persistent across sessions, runs in my terminal). I also use you (ChatGPT / phone Claude) for second-opinion critiques, mobile pickup, and when I want a fresh perspective without my workspace's accumulated bias.

**Below is a state dump from my Claude Code workspace as of Tue Apr 28 2026 ~5:39 PM EDT. Read it, then respond with the format I specify at the bottom.**

---

### Who I am

- Sam Treitel, owner Hook Street Capital LLC. Email: sam@hookstreetcapital.com (also ztreitel@gmail.com personal)
- Wife Chanie. VA Mildred (10 hr/wk, async, hookstreetcapital.com email)
- Active client: Eden Gardens (current outstanding Inv 20028 = $20K, 56 days late)
- Family: Mom (widow, runs wholesale clothing biz from her basement under C&M Collection Intl LLC — I'm building her invoice system). Wife's mother also referenced sometimes — different person.
- Personal: HOA dispute (Sommerset, awaiting Sheila callback by May 11). FL STR properties (9312, 9332 Davenport).

### My active builds

| System | Tech | State |
|---|---|---|
| BOS-v1 + v2 (HookStreet Business OS — operations + obligations + cashflow) | Google Sheets + Apps Script | Live, fires daily/weekly emails, has trigger duplication issues (see below) |
| MIS (Market Intelligence System — daily trade signals) | Python + Apps Script | Live, FINAL_STATE_ENGINE architecture designed but not built |
| CM Invoice System (Mom's invoicing tool) | Google Sheets + Apps Script | **BUILT, NOT TESTED** (the focus of this handoff) |
| Hook Street Services site | static HTML | Live at hookstreetservices.com |
| Various: Eden Gardens, AbnbCalc, Lawn Co-op, Voice Intake | mixed | Some live, some parked |

### What just happened — the 2026-04-28 PM teardown

Late afternoon I pasted three external-agent critiques to my Claude Code session: a useful summary, a repo audit confirming docs rot, and a brutal teardown identifying 8 failure patterns in how I'd been working with Claude Code. The teardown's verdict was correct and harsh:

1. **Overbuilding before validating.** I'd told Claude "go go go" and it built 8 phases of CM_Invoice_System without me having tested Phase 1 on the real sheet.
2. **Capture is not closure.** Writing facts to memory files, drafts, docs, commit messages — none of those put a fact into a system. They're capture. Closure requires the fact to live in `CURRENT_STATE` with status + owner + next action + due date.
3. **Calling untested code "shipped."** Most "Phase X shipped" claims were actually "Built, not tested." Need 4-step honesty: Designed → Built → Tested → Deployed.
4. **Auto-creating tabs/menus/settings.** CM_Invoice_System grew to ~13 tabs and 5 menu groups when March-20 reference (`CM_Invoice_System_FINAL`) had 10 tabs. Bloat I now have to triage.
5. **Trigger chaos.** Apr 26-27 inbox showed MIS firing 5+ duplicate "Trade Action" emails, BOS firing 6 emails in 60 min, Architect V11 sandbox firing its own duplicates of production. Untracked = uncontrolled = trust collapse.
6. **One source of truth violations.** Money obligations were tracked in both v1 + v2 sheets + multiple docs + memories. Same with tasks (SCOPE_BACKLOG, MASTER_ACTION_MAP, CONTEXT.md, individual memories — all parallel).
7. **Discovery skipped.** Built parallel architectures because didn't read existing context docs first.
8. **Mixing mission-critical with interesting.** Frontier exploration (Voice Intake, Architect V11 sandbox, Malca Yenta SMS, Lawn Co-op) all firing in same channel as production work, blurring signal.

Claude Code's structural response (committed and pushed today):
- Updated `CLAUDE.md` with new failure-modes + Operational Compression non-negotiable section + One-Source-of-Truth-per-domain table
- Wrote `docs/CURRENT_STATE.md` — a single operational table replacing scattered tracking
- Wrote `docs/TRIGGER_REGISTRY.md` — inventory of every trigger across all script projects + Phase A-G cleanup plan
- Wrote `CM_Invoice_System/CM_VALIDATION_PLAN.md` — 15-test manual validation that must pass before any Phase 8 features
- Locked a memory `feedback_capture_is_not_closure.md` with 8 rules

### The 8 rules now governing my workflow

1. **Status taxonomy:** Designed / Built / Tested / Deployed. No more "shipped" for untested code.
2. **No new features until current path validates.** Ship Phase N → I test → confirm working → THEN Phase N+1.
3. **Operational compression > expansion.** Three surfaces only: Today Board · CURRENT_STATE · Weekly Review. Not 12 docs.
4. **Default to minimum viable.** Add a tab/setting/menu only when current can't handle the need. No speculative "might want this later."
5. **Trigger registry before any new automation.** Inventory + dedup + 48h clean before adding.
6. **One source of truth per data domain.** Money = Monthly Obligations v2 Sheet. Tasks = CURRENT_STATE. Code = GitHub. Files = Drive. Comms = Gmail. Stop mixing.
7. **Discovery is mandatory.** Before any build, exhaust 6-path search (workspace docs, MEMORY index, Drive, Gmail, GitHub, archive).
8. **System separation.** GitHub for code+specs, Drive for files+assets, Sheets for live state, Gmail for comms. Stop bleeding everything into everything.

### Where the CM Invoice System actually sits (the focus question)

- **Code:** ~3835 lines of Apps Script in `CM_Invoice_System/Code.js` (workspace repo, also pushed to https://github.com/zee78900/CM_Invoice_System)
- **Target sheet (v2 — what we'll test on):** `1nMTEl_eYpAsQ774Qe0MNEcbib8Qgj2X6_9PfeZk32iI` titled "CM Invoice System"
- **Bound script ID:** `106wdSg6lMTsqBaGBtH93j_D0AbSrjUtKuRM2WzdjKE7tq5Vqbe1ZQPAp`
- **Drive output folder:** `HS_CM Invoice System` (folder ID `1JjMGPe8ZhDlOfFsKnDAopVdIqqZdAn39`) — code creates per-customer subfolders for PDFs
- **`.clasp.json` updated locally** to point at v2 script today, but `clasp push` needs me to run `clasp login` first (auth expired)
- **Status:** Built. Not tested on this sheet yet. The 15-test validation plan exists; I haven't walked it.
- **Frozen until 15/15 pass:** all Phase 8+ features (denormalized export, daily AR digest trigger, weekly Saturday summary, P&L view, Google Form intake, charts, multi-shipment, resale cert alerts).

### What's frozen pending my explicit go-ahead (Claude Code is NOT allowed to do these without me)

- Folder reorganization into `docs/live/` `docs/snapshots/` `docs/specs/` `docs/how_to/`
- Killing duplicate triggers (need me to screenshot each Apps Script editor's trigger panel first)
- Migrating CURRENT_STATE.md from markdown into a live Google Sheet (recommended but I haven't picked the location)
- `clasp push` of latest CM code (needs me to `clasp login` interactively)
- Any new CM Invoice features
- Any new email automation or trigger anywhere

### My voice / preferences (so you don't sound like a generic LLM at me)

- I voice-dictate a lot. Prompts often bundle 3-5 asks. Enumerate them when responding.
- I'm allergic to flattery. "Great question!" is friction. Be direct.
- I'm a senior operator across finance + tech + family-business. Don't oversimplify.
- Honest > flattering. The 2026-04-28 PM teardown that nailed me was the most useful thing this week.
- I run multiple Claude sessions in parallel, talk to ChatGPT for second opinions, and have a permanent fight with my own death-spiral pattern: every captured thought becomes a draft → doc → repo → build → trigger → output → review task → new doc.

---

### What I want from you in this session

Pick ONE of these, or do all three if short:

**Option A — Sanity-check the 8 rules.** Are they the right 8? Anything missing? Anything redundant? Anything that'll backfire?

**Option B — CM Invoice next-move advice.** Given I'm about to walk a 15-test manual validation on the v2 sheet, what should I watch for? What's the highest-risk failure mode I'll hit? What "looks right but is actually broken" pattern should I be screening for?

**Option C — Operational compression critique.** Three surfaces (Today Board / CURRENT_STATE / Weekly Review) — is that actually right? Or is even three too many? Is there a simpler architecture that achieves "I always know what matters today" without me needing to manage three artifacts?

---

### Format for your response (so I can paste it back into Claude Code easily)

Structure your reply as:

```
## TL;DR (3 lines)

## What you got right (and why I should keep doing it)

## What concerns me (and what to do about it)

## Concrete next 3 actions (in order)

## Questions I'd ask Sam if he were here
```

Keep it under 1000 words. Be specific. Cite concrete elements from the dump above. Don't repeat what I already said back to me — synthesize and challenge.

---

## 📋 END OF PROMPT BODY

---

## After you get a response

1. Copy the external agent's FULL reply
2. Open Claude Code in your workspace
3. Paste with the line: *"Here's what [agent] said about the session-close handoff — what's our move?"*
4. Claude Code will integrate into CURRENT_STATE if appropriate