Field notes from operators building with Claude Code.
Three ventures on one stack. Mile High Golf. Quantum Caddy. Parley. Field Notes is what I learned that week. The Playbook is the durable reference underneath. No theory. Receipts only.
Bundles you can drop into your repo today.
Every bundle is a self-contained artifact extracted from the live work across QC, MHG, and Parley. Not a screenshot of how I work. The actual files I work from.
Sprint Contract System
Pre-work contracts co-signed by builder and evaluator. Eliminates close-enough shipping. 48 contracts shipped.
- Adapted from Anthropic's Harness 2.0 research
- 6 worked examples from QC, Parley, and the Mile High software
Postmortem Discipline
Blameless postmortems with structural action items. The discipline that makes the same mistake stop happening twice.
- 26+ postmortems shipped across QC and Parley
- Owner-sentiment section is required, not optional
CLAUDE.md Template
The annotated skeleton every Claude Code operator needs before their second week. Three worked examples, one framing essay, and the advice everyone skips: what NOT to put in.
- 3 worked examples: solo founder, 5-person team, agency with client rotation
- Covers the load-everything vs. read-on-demand decision explicitly
Agent Routing Template
The chief-of-staff + N-specialists scaffold. Routing table, authority levels, and an escalation protocol — the piece everyone skips until it bites them.
- 3 worked examples: solo founder, three-venture stack, 5-person team
- Escalation protocol with a full worked example — the most-skipped piece
Vault Structure
13-folder Obsidian template that scales across multiple ventures. The operating system around your AI agents.
- 13 polished templates, ready to use
- Karpathy-style research synthesis layer included
Memory Architecture
The MEMORY.md index pattern. Four memory types, a staleness protocol, and the one rule that stops memory systems from becoming archives.
- Four memory types: user, feedback, project, reference — each with its own shelf life
- Staleness protocol with date-stamp conventions and prune cycle
Agent Hierarchies
Authority levels, the mothership pattern, generation/evaluation separation. The structural decisions most multi-agent setups skip until they break.
- Five authority levels from Observer to CEO — each with distinct write-access rules
- Generation/evaluation separation: why the agent that built it can't grade it
Token Discipline
Context window hygiene. What belongs in CLAUDE.md vs memory vs code. The compaction strategy that keeps sessions fast after month 3.
- CLAUDE.md size target: under 200 lines — if it's longer, information leaked in that belongs elsewhere
- The load-everything vs. read-on-demand decision, with the pointer pattern that solves it
TruPath Stack
The assembled picture — how agents, hooks, memory, sprint contracts, kanban, and the Telegram bridge connect into a single operating system for a multi-venture portfolio.
- Three ventures, five agents, 15 hooks, one Supabase kanban — wired together
- Telegram bridge: phone → Claude headless → portfolio status in 30 seconds
Hooks Bundle
Five production-tested Claude Code hooks — session loading, config protection, eval auditing, decision logging, anchor-check cadence. The mechanical enforcement layer your CLAUDE.md rules have been promising.
- All 5 hooks read real Claude Code JSON input — none is pseudocode
- Config-protection hook prevents the class of incident that cost 4.5 hours of test setup
The Security & IP Discipline Playbook
Five gates that keep a fast-moving AI shop from leaking its secrets, its source license, or its patent rights. Written the day I pasted my session tokens into a chat.
- Disclosure, license, secrets, agent-code, and IP-review gates — each with a checklist
- Grounded in the 2025–2026 record: the GitHub key leaks, the AGPL forks, the trade-secret cases
The CV-for-Physical-Sports Wedge Playbook
Don't build computer vision for sports. Build it for one sport — the one with the most constrained geometry and the simplest rules — get it referee-grade, then generalize. Three filters and the trust-bar framework.
- Three filters that turn 'CV for sports' into a winnable first product
- The trust-bar framework — why 96% accurate is a dispute generator
Why run a parallel research arm
A small operator can run a parallel research project alongside two startups without it cannibalizing them — but only under specific conditions, and Parley is the worked example.
- The decompression-channel + underserved-research-area thesis
- Three conditions that mean don't start one
How to run one without it cannibalizing the startup
Four discipline levers that let a side research arm survive contact with the main work — scope rails, monthly cadence, public publishing, and decision logs, with kill criteria written down in advance.
- Scope rails + monthly cadence: the two levers that absorb the main work's variance
- A public publishing loop + decision logs: borrowed deadlines + a real audit trail
The Honest Sign-Language Research Playbook
The discipline behind Parley's notebooks: question-first contracts, signer-holdout splits, multi-seed floors, failure-modes-first, and a Deaf-community honesty checklist. The rules that make a 45% you can trust beat an 85% you can't.
- Pre-registered hypotheses + signer-holdout eval, from real notebooks
- Report the per-signer distribution, not the pooled mean
Pre-register the recipe before you rank architectures
A shared training recipe across architectures does not control for the recipe. It bakes it into the result and disguises it as an architecture finding. Four cheap controls keep the ranking honest.
- Why a shared recipe measures the recipe, not the model
- The four controls: per-architecture recipe, one hardware, seed spread, shuffled-label check
Latest Field Notes
I called three sign-recognition models failures. The recipe was the failure.
A Parley notebook reported three landmark architectures as broken on cross-signer ASL. A warmup and a gradient clip brought all three back, and two matched the best model. The ranking had measured my training recipe, not the models.
Recipe beats architecture: lottery tickets in sign models
We trained seven landmark architectures three times each. Three of them worked on one seed and collapsed to near-random on the others. A single-seed comparison would have called two of those collapses a result.
The 38-point gap: one accuracy number, twenty-one very different users
Our sign model averages 42% across signers. That average hides a range from 26% to 64% — and the thing that decides where a person lands is not the signs they make, it is who they are.
How the operator loop runs.
Capture turns into knowledge. Knowledge turns into shipped work. Shipped work that breaks turns into postmortems. Postmortems turn into rules. Rules carry forward to the next session. The compounding part is the actual product — the rest is just the mechanics that produce it.
Every chat, every meeting, every observation, every postmortem note. Drop into 00-Inbox/ or 09-Research/Inbox/ first. Don't try to organize at capture time — just don't lose it.
The Vault Structure →The Playbook
The lab
Pre-launch. SBA underwriting in progress. Operations management runs on Claude Code agents.
Phase 0 MVP. Smart-board hardware + CV pipeline + ESP32 firmware. The engineering testbed for everything we write about.
Phase 0 research arm. Sign-language recognition. Hard reasoning problems, in the open on Kaggle.