Post Title
The Claude Conductor: how one question led to an AI review of my AI workflow
Series (optional)
Claude Code Chronicles
Estimated publish
2026-Q2 (imminent — draft complete)
Tags
claude-code, methodology, stalactite-principle, clone-force-99, typescript, bad-batch
Summary
A conversation that started with a simple question — can Codex make Claude better at TypeScript? — turned into something unexpected: an outside model analysing the entire Bad Batch workflow and reflecting it back. Not a critique. A mirror. What it named correctly, what it surfaced that we'd stopped looking for, and what evolved as a result. The thesis: everyone tells you to double-check AI responses. Nobody tells you to build a system where double-checking is structural. I did. I automated it. It's called Clone Force 99.
Sections
One question
The frustration that started it — Claude Code reaching for as any when it loses confidence in a TypeScript pattern. Someone online arguing Codex was the answer. Going to test that claim and ending up somewhere unexpected.
The system, reflected back
What happened when an outside model had to analyse the workflow to answer the question. The things it named that had been working intuitively but never formally articulated — artifact-centric over conversation-centric, reconstructable intelligence from artifacts, personality as architecture.
What we'd stopped looking for
When a system works well you stop auditing it. What the outside perspective shook loose: duration tracking from GitHub label transitions, session intent lock as a first-class pattern. Both were latent in the system. The right questions surfaced them.
The real answer to the TypeScript problem
Not a different model — a stricter mission contract. When Crosshair owns a fix with explicit constraints baked in, the escape hatch isn't available. The system enforces discipline the model won't enforce alone. That's the whole point of the clone specialisation layer.
On authorship
The stalactite doesn't have one author. Claude Code, Claude Chat, and the developer built this across hundreds of sessions. That's not a disclaimer about AI doing the work — it's the Stalactite Principle in action.
What's next
The open problem: compression. What happens when the stalactite gets too heavy to reason about in a single context window. Part 5 is about that ceiling.
Key Takeaways
- When your workflow feels solid, that's exactly when an outside perspective is most valuable — you've stopped looking for what could be better
- Personality in Claude Code isn't cosmetic — named clones with defined specialisations constrain behaviour more reliably than instructions alone
- The answer to AI making mistakes isn't double-checking manually — it's building a system where mistakes have nowhere to hide
- Duration tracking from GitHub label transitions gives you real lifecycle data across sessions, not just within them
- The stalactite has no single author — that's not a disclaimer, that's the point
Post Title
The Claude Conductor: how one question led to an AI review of my AI workflow
Series (optional)
Claude Code Chronicles
Estimated publish
2026-Q2 (imminent — draft complete)
Tags
claude-code, methodology, stalactite-principle, clone-force-99, typescript, bad-batch
Summary
A conversation that started with a simple question — can Codex make Claude better at TypeScript? — turned into something unexpected: an outside model analysing the entire Bad Batch workflow and reflecting it back. Not a critique. A mirror. What it named correctly, what it surfaced that we'd stopped looking for, and what evolved as a result. The thesis: everyone tells you to double-check AI responses. Nobody tells you to build a system where double-checking is structural. I did. I automated it. It's called Clone Force 99.
Sections
One question
The frustration that started it — Claude Code reaching for
as anywhen it loses confidence in a TypeScript pattern. Someone online arguing Codex was the answer. Going to test that claim and ending up somewhere unexpected.The system, reflected back
What happened when an outside model had to analyse the workflow to answer the question. The things it named that had been working intuitively but never formally articulated — artifact-centric over conversation-centric, reconstructable intelligence from artifacts, personality as architecture.
What we'd stopped looking for
When a system works well you stop auditing it. What the outside perspective shook loose: duration tracking from GitHub label transitions, session intent lock as a first-class pattern. Both were latent in the system. The right questions surfaced them.
The real answer to the TypeScript problem
Not a different model — a stricter mission contract. When Crosshair owns a fix with explicit constraints baked in, the escape hatch isn't available. The system enforces discipline the model won't enforce alone. That's the whole point of the clone specialisation layer.
On authorship
The stalactite doesn't have one author. Claude Code, Claude Chat, and the developer built this across hundreds of sessions. That's not a disclaimer about AI doing the work — it's the Stalactite Principle in action.
What's next
The open problem: compression. What happens when the stalactite gets too heavy to reason about in a single context window. Part 5 is about that ceiling.
Key Takeaways