The Problem
I wanted to know whether two open-source tools were worth wiring into my setup: hands-on-deck from EveryInc, and visual-explainer from nicobailon. hands-on-deck claimed to let an AI agent edit PowerPoint files precisely — patching the existing file instead of regenerating it from scratch. If that worked, it solved a real headache: keeping a branded deck on-brand while an agent does the work. I had an actual deck waiting that needed both content edits and a visual lift before a high-stakes external briefing, so I pointed it at that instead of a toy.The Approach
I had Claude install both tools and gave them non-overlapping lanes so I’d always know which one produced a given artifact. Then the real test: rebuild the briefing deck. I kept the redesign in a separate file from the start — whatever the experiment did, the version that had to ship was protected. The design decision that mattered was upstream of any slide. Rather than ask for a generic “make it look better” pass, I had Claude pull the company’s existing sales deck apart and extract its actual design system — the type pairing, the color rhythm, the recurring visual devices — so the redesign would be built from the brand’s own grammar. One small prerequisite turned out to be load-bearing: installing the brand’s fonts locally. Without them, the agent’s renders were quietly substituting fonts and weren’t showing me the truth.What Happened
Claude did the building — the reorder, the content rebuild, then all fifteen slides authored in HTML and compiled back to PowerPoint. Its renders looked clean. But the deliverable lived in a different slide tool, and the only honest test was a full round-trip: export, re-import, look. The round-trip was less forgiving. The import surfaced four failures the agent’s own renders never showed — highlight boxes that vanished when they sat mid-sentence, a serif headline that rendered taller on import and collided with the text beneath it, a customer quote whose text silently dropped, and faint shadows the agent insisted weren’t in the file. It was right that they weren’t in the file; the slide software was adding them on import. Claude had to go digging through the raw XML to prove where they came from (to its credit, it did). Two round-trip cycles later it imported clean. I caught one thing it missed on my own — the title had picked up an unnecessary bold weight — and took it off myself.What I Learned
The agent’s self-verification is only as good as its rendering environment, and that environment was lying to it. The render is not ground truth; the round-trip is. Almost none of the craft here was in the building — Claude did that — it was in insisting on the real test, and in holding “it looks right in my render” and “it survives the round-trip” as two genuinely different claims. Both tools are staying in the setup, and the round-trip is a standing step now.Drafted by my Claude from our working session. I approved it before it went up.