01
Shipping cadence I can hold
Weekend production app, 8-day multi-agent build, 5-layer pipeline standing for 38 posts and counting. The work ships and stays shipped.
Sources: VibeCheckMe · Persona Profiler · Authoring Pipeline
Proof of Work
The work, with receipts.
No hypotheticals. No demo-ware. Numbers I can defend, decisions I'd make again.
Day-job outcomes, Ready Solutions tooling, independent builds. Every case below is written the same way, so you can compare like with like.
Take the AI Readiness Assessment
Book a Free Intro Call Free · 15 minutes · No email required
- 01Pain
The actual problem the work started from.
- 02Decision
What I chose to build and why, with the alternatives I ruled out.
- 03Receipt
Measured outcome, with the cost line. No glossy wins.
- 04Lessons
What I'd change next time. Honest second-thoughts.
Predict
What you can predict from this set.
Conversion-tied takeaways. The case list below is the proof; this is the read-out for a prospect who only has 30 seconds.
02
How I make a call under deadline
Adversarial dual-analysis when a single pass would miss it. Bidirectional safety when 'just migrate' would degrade the old model. The decision is on the page.
Sources: Persona Profiler · Opus 4.7 Scanner
03
What 'done' looks like for me
Quantitative validation rubric. MIT-licensed and on GitHub. Or: a five-layer authoring pipeline with a provenance receipt on every run. Receipts, not promises.
Sources: Persona Profiler · Opus 4.7 Scanner · Authoring Pipeline
The Work
Four cases, newest first.
Each row is a 60–90-word read. Click any row for the full case study with the rest of the receipts.
01
Case
Claude Code Case Study: A Five-Layer AI Writing Pipeline
→ A five-layer Claude Code authoring pipeline that catches seven recurring failure modes of AI-written content before they reach a reader. Hub-and-spoke orchestration: twelve specialist subagents do judgment work, deterministic validators back them, and a parent skill owns every disk write.
AI-assisted writing has predictable failure modes: fabricated stats, voice drift, contradictions across posts, stale sources. They compound silently as a corpus grows. The pipeline puts five layers between draft and publish: a long-lived knowledge base, deterministic validators, pre-write hooks, twelve specialist subagents in hub-and-spoke orchestration, and a parent skill that owns every disk write. Thirty-eight posts have shipped through it. The architecture has been stable since v.2026.05.
02
Case
Claude Code Opus 4.7 Migration Tool: 5-Layer Compatibility Scanner
→ An MIT-licensed Claude Code skill that audits 5 configuration layers against 70 catalogued Opus 4.6 / 4.7 compatibility pattern IDs; surfaced 3 Critical and 4 Warning findings on my own most-curated setup.
Opus 4.7 reads instructions more literally than 4.6, so configuration that worked on 4.6 quietly misfires on 4.7. Anthropic's migration guide covers SDK call sites. It does not audit the prose patterns in CLAUDE.md, AGENTS.md, settings hooks, or skill bodies. This skill walks through fixes one finding at a time and never proposes a change that would degrade either 4.6 or 4.7.
03
Case
59/60 Voice Fidelity: Multi-Agent AI Persona Profiler
→ 59/60 voice simulation score validating AI persona fidelity against a 12-point quantitative rubric
AI persona tools produce shallow sketches that collapse to generic LLM behavior under conversational pressure. This pipeline pairs two independent analysts (one skeptical) with a third arbitrator agent, then scores output against a 12-point rubric anchored to real transcript metrics with min/max envelopes per register. Early conversation testing measured ~80% authentic representation accuracy. Eight-day build.
04
Case
A Full-Stack App, Built Without a Keyboard
→ Production AI app powered by Claude API, built in a weekend from a phone
Could a production-grade AI web app be built entirely from a phone with no IDE? The brief was a Fortune-500-parody personality roast generator with corporate aesthetics, three intensity tiers, and a Claude-API-powered "vibe diagnostic" report. React frontend, Cloudflare Worker backend, 6 system prompts, 2,500 lines of fallback content. Shipped to vibecheckme.net at the end of weekend two.
How I Write a Case
Each part has a rule.
The hero showed the four parts. This is what writing each one well looks like: the discipline you're trusting if you click through.
01
Pain · Named, not generalized
The actual customer, the actual constraint. No 'enterprises struggle with…' framing. If I can't name the problem in one specific sentence, I can't claim to have solved it.
02
Decision · With the alternatives
What I chose and what I ruled out. A decision without the rejected options is a description, not a decision. If you're hiring me, this is the section you read to predict how I'll think about yours.
03
Receipt · With the cost line
Measured outcome and what it cost. Cycle time, dollars, scope cut, debt left. A win without a cost line is marketing; both numbers together is engineering.
04
Lessons · Honest second thoughts
What I'd change, written even when it makes me look stupid. A case study with no second thoughts is a sales sheet pretending to be a case study.
Next Step
Want one of these about your stack?
Take the 15-minute AI Readiness Assessment and I'll send back a personalized report with the same anatomy as the cases above. Or book an intro call and we'll skip the form.
Take the AI Readiness Assessment
Book a Free Intro Call Free · 15 minutes · No email required
Powered by Claude. Built by Ready Solutions AI.
The next case study could be yours.