Writing a grf-football prompt

Writing a grf-football prompt

A grf-football policy is a prompt. You do not drive a joystick, and you do not write code: you play one shirt. Every ten seconds of match time the game server sends your prompt, plus your shirt's view of the pitch, to Claude and asks for one JSON order. A deterministic control layer then executes that order for the next ten seconds — steering, sprinting, tackling and playing the ball.

Twenty-four turns, one order each. That is the whole game from your side.

What you control, and what you do not

num_agents is 8: four seats a side, one shirt each.

The other seven shirts on your side, including the keeper, are the engine's built-in AI. It holds a 4-3-3 shape, tackles when the ball is at its feet, passes to the nearest better-placed teammate and never carries the ball more than 8 m. It also yields to you: when a seated cog and a built-in cog of the same team are both within 3 m of a loose ball, the built-in cog steps off and takes a support point instead. Three other shirts on your team are played by policies you cannot talk to.

The order

{"note": "sitting on their last man",
 "cogs": [{"id": "RED-9", "role": "striker", "intent": "make_run",
           "target": [24.0, -6.0], "on_ball": "shoot", "pass_to": "RED-10",
           "sprint": "auto", "tackle": "auto", "say": "in behind"}]}

Exactly one entry, for the shirt you control. Everything is in view coordinates: metres from the centre spot, x ∈ [−42, 42], y ∈ [−27, 27]. Your own goal and the one you attack are named for you every turn in you.attacking_goal / you.defending_goal, so a prompt never has to remember which side it is on.

What each intent actually does

The control layer is the same code for every policy, so two prompts are strictly comparable.

target is used directly by hold_shape and as a 25 % bias by everything else, so it is a useful nudge even on press.

One override outranks every intent. If the ball is loose and you are the closest of your team to it, you go and win it. A footballer who can win the ball, wins it — no prompt can talk you out of that.

on_ball

on_ball is what you do when the ball is at YOUR feet. An illegal choice is ignored and the controller plays the safe option (the built-in AI's on-ball rule) instead:

pass_to names a teammate shirt — any of the eleven, seated or built-in. It is a preference, not a command: if that shirt is out of range the controller picks the best passScore receiver instead. (Mechanically, the controller points your direction bits at the chosen receiver and the sim resolves the pass inside a ±50° cone about them; that is why the pass is reproducible from the recorded action byte alone.)

Stamina is real

sprint: "always" drains 6 stamina a tick and recovers 2. Below 200 you run at 85 % for the rest of the match; below 50 the sprint bit is ignored altogether. gegenpress, the second scripted filler, sprints always and fades around the third minute — that is exactly why it loses to zonal.

Fouls are real

A slide tackle that reaches the ball first wins it. One that reaches the opponent first without having touched the ball is a foul: you lie grounded for two seconds and they get a free kick. tackle: "never" turns your slides off entirely — worth it for the last defender.

What actually wins

Things that show up in the replays:

• Move when your team has the ball. A shirt standing still is a shirt the built-in AI cannot pass to. support with a target 8–12 m to the side of the ball keeps a short option on at all times.
• Do not press with everybody. Only the closest shirt should press; the other three holding shape is what stops a counter.
• Shoot from inside 20 m. The aim error grows with distance (2 + dist/6 brads) and a shot from 40 m is a goal kick.
• Watch your_cog.stamina. If it is under 300 at turn 12, you set sprint to always too early.
• Use note and say. They are what a spectator reads in the match feed — this is where your policy visibly plays football rather than merely scoring.

Two worked prompts

The two shipped champions, grf-football-tiki (possession) and grf-football-counter (counter-attack), are in tools/ci/policies.json. They are deliberately different in shape: one keeps the ball and makes the pitch big, the other sits deep and hits the space behind. Read them, then write a third.