What's actually in my CLAUDE.md (and what I deleted).
The soul file is the highest-leverage document in the whole stack. Most of its value is deletions.
What happened, what we learned, and what changed. Stories from the space between an idea and something that works.
The soul file is the highest-leverage document in the whole stack. Most of its value is deletions.
Most postmortems produce promises. Ours produce mechanical rules. The template is the difference.
One page, seven rows. The difference between an agent that ships and an agent that improvises the definition of done.
683 impressions, 6 clicks. Five title rewrites, one diagnosis: every title promised a format instead of an answer.
A more capable coding agent can cross a larger problem in one run. That makes scope, receipts, and independent verification more important, not less.
A knowledge graph becomes another inbox unless it is clear what belongs there, what stays elsewhere, and which system may call work complete.
Terrain, trees, bunkers, and mountains were not enough. Technical completeness and visual authorship are different jobs.
Before there was a playable hole, there had to be a way to keep concepts, assets, physics, builds, and claims from collapsing into one folder called progress.
The Mac app was telling the truth about paired capture while hiding the working single-camera path below the preview.
Moving a golf video off the iPhone was easy. Proving that the phone copy could be safely removed was the real system.
Swing Trainer's Live Studio passed its presentation suite, but a real 1149×772 window clipped capture controls. Visual checks stay a separate release gate.
A signed macOS process, green tests, and an idle loop did not prove a usable app. The investigation separated disk work, window state, and observation error.
Swing Trainer had to keep partial human review without calling incomplete work training truth. Drafts, labels, and source video stay separate.
Body motion can start and stop a golf recording, but it cannot prove where the club is. The evidence boundary behind Swing Trainer's hands-free capture system.
The physics boundary inside Swing Trainer: what monocular video can support, what calibration might unlock, and what still requires an external instrument.
Why the Swing Trainer RunPod gate stops at provenance, frozen splits, and real five-point labels—even with thousands of images already staged.
GitHub, Roboflow, Kaggle, and Hugging Face produced plenty of files. The hard part was refusing to call mismatched labels a dataset.
The club looked simple until motion blur, hands, clothing, and the edge of the frame turned one object into five different evidence problems.
The useful question is not “can Airtable automate this?” It is “what may this automation change without a person reviewing it?”
A pose system sees a body, a club detector sees a club, and a launch monitor reports a ball. Treating all three as one measurement invents facts.
The Range spike has a full instrument, a protocol, and a hard exit bar. It still has no iPhone measurement. That is not a gap in the story. It is the story.
No measured data, only a simulator to trust. A model-vs-model audit still produced decision-grade findings: the five moves, and what the method cannot tell you.
Most sessions return confident prose. This one returned reproduction scripts, error bars, and ranked experiments. The brief was written like a lab contract.
A tens-of-millions run, and one nearly three times larger. They agree on every trajectory, disagree on half the outcomes, and every max score was an artifact.
Why a markdown vault outperforms a Notion plus Asana plus Drive plus Slack stack when AI agents are part of the work. Battle-tested across QC, MHG, and Parley.
Twenty-six postmortems across QC and Parley in six weeks. The template, four examples, and how to stop the same mistake from happening twice.
A 48-contract system born from agents that praised their own work. The fix wasn't a smarter model. It was structural.
Architectures that looked broken on cross-signer ASL came back after a warmup and a gradient clip. Two matched the best. The ranking had measured the recipe.
Seven landmark architectures, trained three times each. Three worked on one seed and collapsed on the others. One seed would have called two collapses a result.
Our sign model averages 42% across signers, hiding a range from 26% to 64%. Where a person lands depends on who they are, not the signs they make.
Our best landmark-only sign model scores 45% on signers it has never seen. The field reports numbers twice that high. We publish the lower, honest number.
A running catalog of how hearing-led sign-language AI fails. Not a list of other people's sins. It exists so Parley can catch itself doing the same.
Field notes from an operator running three ventures on Claude Code. Biweekly. No theory. Receipts only.
One AR-glasses decision for two ventures. The criteria that survived the cut, and why doing it once was the right move.
I started a Kaggle research project in Q2 2026 while running two startups. The decompression channel, the four open questions, and what makes it survive.
Claude writes ESP32 firmware that compiles, but cannot reliably predict what it does on the hardware in front of you. Three incidents, and the gate I added.
Systematic grid search gives you coverage. Random sampling gives you reality. You need both, and you need to know what each one is telling you.
A parametric physics engine, 8 parameters, two overnight runs. An honest account of what building it took, and what we could not have learned any other way.
AR on physical objects is 80% coordinate system problems. Claude is great at helping you think through the geometry. You still have to understand it yourself.
Clean data beats model size. Every time. Don't upgrade the model until you've audited the labels.
LLMs in real-time hardware aren't ChatGPT. Latency budget is the constraint that changes everything.
Agents are unreliable judges of their own work. Here's how a structural fix — not a smarter model — stopped QC's CV pipeline from shipping silent failures.
Sixteen cookies that together are my whole Google account, dropped into a chat window — while building a security playbook. The how is the whole point.
Twelve-camera rigs and Hawk-Eye sit at the top. The same capability now fits on a $249 board and a commodity camera, and the smartest way in is the narrowest.
The routing model that holds up under load. Built for TruPath, tested on Mile High Golf, Quantum Caddy, and Parley.
A 20-agent Electron app that didn't ship. The decision to throw it away. What I'd tell anyone tempted to build the same thing.