Jason Poindexter: Design director who ships production AI

Eighteen months shipping production AI with Claude Code. Five seams I keep hitting.

I am a designer who codes well. Four production AI builds in eighteen months, all of them shipped with Claude Code as the daily tool. Monoth. FABRK. Gripe. Interrogation. I am not the user Claude Code was built for in 2024. I am close to the user it is becoming. Notes below come from real shipping work, not commentary from outside.

1. Long runs collapse the moment context drops

A long task starts well. Twelve files in, the agent loses the thread on a constraint I gave it ninety messages ago. The output is still confident. That is the problem. I want a soft pin. A constraint promoted to the top of the run, visible while the model works. At the end, a small reconciliation, I held these, I drifted on these, reconsider these. That last bit is the trust move.

2. Terminal output is a typographic problem

A multi-file diff, a directory listing, a search sweep, and a paragraph of reasoning all share the same scroll. The user does the parsing the system should do. I want three visual registers. Prose, code, structural data. Different leading. Different density. Diffs collapsed by default, expand selectively. Search results that promote the matched line and dim the rest. Terminal as substrate is fine. The visual grammar inside it has not caught up to what the model outputs.

3. Trust signals at the edit moment

The agent says it will edit a file. There is a half second to intervene or trust. Right now the choice is implicit. Make it explicit when risk is real. Editing a string, zero friction. Editing a migration, an env file, a security boundary, a soft confirm with the preview rendered. Not to slow the agent down. To make trust legible.

4. Multi-file plans need preview before commit

Anything spanning more than two files needs a plan I can scan before edits land. Not a dry-run flag I have to remember. A default plan view, change set as a tree, each leaf with a one-line intent. Approve the whole plan, approve a subset, reject and explain. Today the model asks, I say yes, edits stream. Fine for small tasks. Breaks on the refactors that are exactly when the agent earns its keep.

5. Prototype mode and production mode are not the same product

Prototype tolerates everything. Hardcoded data fine. Vibes fine. Second pass fine. Production tolerates none of it. The product does not visibly distinguish the two. I want a posture switch. Not a setting. A state the agent enters when I ask it to ship. Stricter checks. Mandatory verification. Different default on tests. Different posture on confidence. The model already behaves this way when prompted. Make it a first-class affordance and adoption changes.

Why this matters