Most features are prompts once you expose the data

The argument. The speaker says prompt engineering is overrated, not because prompts should be sloppy, but because engineers over-specify. They think in code, so they turn every request into an implementation plan. The useful move is to ask the agent for the outcome and let it work out the steps. That sounds simple, yet it changes what gets designed.

The evidence. He shows how asking instead of building surfaced real data. A friend wanted a T3 Code feature to show which tool calls are slow. It already existed. Claude coding failures traced to ripgrep syntax that worked in bash but not zsh. A home directory search took 41.7 minutes. A context visualizer showed Railway was injecting 70,000 tokens into every request. None of this required a new feature. It was a prompt over existing data.

The pattern. The underlying shift is from building features to asking for information. Tool-call timing, context usage, and capability descriptions all live inside the agent's world. Engineers rarely ask because they assume a feature request is needed. Once you frame it as a query, the same data becomes available on demand, and then a chat can turn it into a dashboard, a report, or an action. This is the mental model most teams still lack.

The redesign. The T3 Triage example is the strongest part. The first version was a complex flow with branches and hooks. He replaced it with five lines of code and 30 lines of markdown. The prompt tells the agent to ask what went wrong, read machine facts, check a playbook, inspect source, and file an issue. A support flow became a prompt with a playbook. That is a concrete pattern teams can copy.

The tradeoff. Prompt-first features work when failure is acceptable and speed matters. They do not give deterministic UX, role-based permissions, audit logs, or reliable SLAs. The video mostly ignores those costs, plus token spend and latency. The 70k token example shows prompts can also hide expensive waste. Teams still need primitives, data pipes, and MCP interfaces before prompts can replace something real.

The vendor line. The video is framed around T3 Code and includes a sponsor segment. Claims about feeling like more of an engineer than ever are personal and unproven. The 'half of features should be markdown' estimate is provocative, not measured. Engineering leaders should take the mindset, not the product positioning, and validate it on their own backlog.