Masking slides for AI: placeholders that survive the model but break the layout
shoumikgoswami · reddit · 2026-10-27
Building OpenGamma with Claude Code, the author replaced sensitive client values with placeholders before letting the AI work on decks, then restored the real values at PowerPoint export. The trap: [CLIENT1] is short, "Northwind Logistics Holdings" is not — slides designed around the token look fine but break when the full name returns.
Key lessons:
- Size text for the real value, not the placeholder; the same applies to table cells and chart labels.
- Two separate checks are needed: did the sensitive value stay out of the model's view, and does the finished deck still work after restoration? Passing one doesn't imply the other.
Design principle: don't hide a value and accidentally hide everything the next step needs to work.
More from coding & agent
- Context Compaction Destroys Provenance in AI Agents — How to Fix It — arpit_bhayani · 2026-10-27
- Polylane builds its agent memory system on Git, with nightly self-correction — dinasaur_404 · 2026-10-27
- Manus on rebuilding itself as an AI-native org: expose interfaces, demand evidence, go async — dotey · 2026-10-27
- A peek at a personal multi-agent setup: Codex, Muse and Grokbot with role-based memory — op7418 · 2026-10-27
- Stanford's BioBabel: AI Agents Auto-Port Scientific Packages Across Languages — anshulkundaje · 2026-10-27
- Context Layer V2 evals: ~40% accuracy gain, ~30% latency cut vs MCP alone — shensi · 2026-10-27