Clauding with Jenn.

One Command, Five Trust Checks

The first Jenn OS replay brief turns the session protocol into a public product story: one closeout command that surfaces worktree drift, runtime collisions, dirty repos, build-log coverage, and deployment nudges in a single operator-readable pass.

The App May Not Be the Problem

The second Jenn OS replay package is a tighter debugging clip: one tiny checker that proves bizarre local behavior is often environment collision rather than bad application code.

When the Warning Is Wrong but the Save Is Real

Task Command showed a cloud-sync warning even though a deleted item stayed gone after refresh. The failure was not the task board alone. It was the proof model around the task board.

Provenance Theater: When the Safety Net Is Made of Paper

Six data explorers had provenance labels, round-number detectors, export guards, and a full DATA_VALIDATOR. Every CFPB complaint count was still fabricated. The infrastructure for rigor is not rigor. Calling the API is rigor.

If a Number Is on the Page, It Needs a Source

My portfolio and work pages had believable but conflicting project counts. The fix was to stop treating metrics like copy and start treating them like derived data with generators, shared imports, and an audit surface.

Jenn OS v1: A Local Operating Layer for Work

The public architecture thread now resolves into a concrete product direction. Jenn OS should start as a local operating layer: one narrow loop, explicit records, package-first wrappers, and a UI that compiles from memory instead of pretending to be memory.

Tag It, Badge It, Block the Export

Six standalone data explorers shipped with embedded reference data that could be exported straight into client deliverables. The fix was a three-part system: tag every data point with its source at ingest, show provenance badges in the UI, and block exports until the data is verified.

Run the Numbers — Bayesian Forecasting as a Mandate

An AI operator kept picking the wrong recovery path under pressure — rebooting the Mac when killing a sprawl of browser workers would have saved 9 minutes and $0 in blast radius. The fix was not more rules. The fix was a forecaster persona whose only job is to price every fork with a base rate, explicit prior, posterior with a range, and named update triggers — before the system acts.

Mar 2026

Compacting a Site Without CSS Zoom

Some pages were only comfortable when I zoomed my browser out. The right fix was not literal zoom. It was a shared shell that made app-like pages calmer, denser, and less shouty on desktop while leaving editorial pages alone.

A Working Protocol for Claude and Codex

The problem was not that Claude and Codex were both useful. The problem was that they could not learn from each other. The fix was a simple protocol for shared memory, canonical files, and promotion of lessons into reusable rules.

Enforcement Over Knowledge

78% of the bugs Claude shipped had rules that should have caught them. The problem was never that the agent didn’t know — it’s that nothing forced it to check.

Six Data Explorers from Scratch

FTI economists were spending hours navigating BLS and Census websites to pull public data for background research. I built six single-file tools that do the same thing in seconds.

Trial-Ready Excel from Code

Consulting deliverables end up in Excel. The quality bar for expert testimony is "would this survive cross-examination?" So I built a Python engine that generates investment-bank quality workbooks.

The Sync Bug That Looked Fine

My health dashboard showed stale data for weeks. It looked fine — numbers were there, charts rendered. The bug was invisible because the data was real data. Just the wrong real data.

When Auto-Commits Obscure History

I had an auto-commit hook that committed after every Claude Code session. 122 commits later, git log was useless — every message said "auto-sync session data."

Feb 2026

Cross-Model Interlocutor

I got tired of catching my own blind spots. So I built a tool that lets Claude ask ChatGPT to poke holes in its work.

Skills as Loadable Context

My first real AI project bombed because I delivered percentages instead of recommendations. Now I write down what went wrong and load it into every future build.

The CLAUDE.md Hierarchy

I run two Next.js sites with completely different color palettes. Early on, Claude kept painting my personal site in corporate blue. The fix was embarrassingly simple.

Plaud Multi-Note Sync

I record everything — meetings, therapy sessions, the stuff I think about on walks. Plaud’s default export missed half the notes, so I wrote my own sync.

Making AI Sound Like Me

I fed Claude my journal entries, my Plaud transcripts, the way I actually talk. Then I told it what never to write. The second part matters more.

Validating AI Output

Claude produces confident, clean output. That's the problem. You can't tell by looking whether it's right. I built a validation methodology that catches errors before they ship.

Tools That Solve One Problem Well

I keep building small, focused tools instead of big platforms. A color contrast checker. A financial model generator. A prompt library. Each one does one thing and actually gets used.

Tariff Exposure — Full System Test

I loaded all four skills, pulled live tariff data, invented a skeptical CFO, and had ChatGPT tear the whole thing apart. The counterfactual is the proof.

The CSS Bug in 51 Files

A single CSS shorthand rule in globals.css silently broke spacing across 51 files. Tailwind’s `pt-12` was being overridden by `padding: 0 1.5rem` — and it took a full site audit to find it.