Spec Writer
Press Begin to receive your first spec.
Empty policy. The agent will do literally nothing. Add instructions below.
How Spec Writer Works
You don't play the agent. You specify it. You assemble a short policy out of instruction blocks, hit Run, and the agent executes your spec literally, one step at a time. It does exactly what you wrote — never what you meant. Every bug is a misinterpretation you authored yourself.
- Tap blocks in the Instruction Set to append them to Your Policy.
- Tap an instruction in your policy to delete it; use the arrows to reorder. Tap the number on REPEAT to change its count.
- Run executes the whole spec; Step advances one instruction at a time; Reset rewinds the agent.
- Guide the agent to the goal. Grab any data chips. Avoid walls and gradient pits.
The Instruction Set
STEP moves one cell forward. TURN LEFT / RIGHT rotates in place. GRAB picks up a chip on the current cell. IF WALL? / IF CLEAR? run the very next instruction only if the cell ahead is (or isn't) blocked. REPEAT n … END loops everything between them n times — your first taste of an actual control-flow bug.
Scoring
Your score is the number of specs solved. Later levels hand out a bonus for shorter policies — fewer instructions, less surface area for the agent to take you literally. Your best run persists in this browser.
Slop Fact: This is the entire field of AI alignment, compressed into a tiny grid. The agent is perfectly obedient and perfectly literal, which is exactly why it walks confidently off the edge into a gradient pit. Specification is the hard part. The optimizer was never the problem.