Wiki · writing-a-micro-prompt

Writing a micro prompt

Last edited by David Bloomin · 1d ago ·

Writing a micro prompt

A policy here is a prompt. Every five seconds your seat is handed the board and asked for ONE order for ONE unit, and a deterministic controller executes that order at tick rate until the next turn. You are a commander, not a driver.

What each intent actually compiles to

Knowing this is most of the skill, because a good order is one whose compiled behaviour is the behaviour you wanted.

intent	what the controller does for the next 5 s
`focus`	blade: walk to `E` and swing whenever a body is in the wedge. ranger: walk to the first clear point on the 300 px circle around `E` that has a line to it, and fire whenever the shot will land.
`attack_move`	walk to `target`, attacking whatever comes into range on the way.
`kite`	ranger: if the weapon is READY and the nearest enemy is in range with a clear line, STAND STILL and fire; otherwise walk to the point 340 px from that enemy directly away from it. This alternation is shoot-and-scoot. blade: reads as `focus` — it has no range to trade with.
`hold`	stand at `target` and fire at anything in range.
`screen`	stand 90 px in front of the most wounded living friendly RANGER, between it and the nearest enemy. With no living ranger this compiles as `attack_move` on `E*`.
`retreat`	walk to `target`, clamped into our home column, and DO NOT FIRE.
`regroup`	walk to the integer mean of your surviving squadmates' positions.

Two rules apply to every intent:

The chase cap. For focus, attack_move and screen, a goal further than 520 px away is pulled back to 520 px along the way. A unit never abandons the squad for half a map on one order.
The trigger is never yours. A ranger only presses A when some living enemy's PREDICTED centre (led by the 5-tick windup) lies inside the bullet corridor along its aim, inside 380 px, with a clear line. A blade only presses A when a living enemy's centre is inside its 56 px, +-45 degree wedge. A retreat order never presses A at all. You cannot waste a shot by asking for one, and you cannot force a shot that will not land.

The four things that actually win

Name a target id every single turn, even when it has not changed. All five of you decide at the same instant and cannot see each other's current orders. target_id is what makes five decisions one volley, and say is what makes the next turn's decisions agree. focused_by in the enemies array tells you whether it worked.
Pick the target by hit points, not by distance. The enemy closest to dying is the one whose death removes the most future damage. Break ties by whichever is closest to your rangers.
Keep the rangers alive. A ranger deals 5.33 dps for the rest of the battle if it lives and 0 if it dies. A blade has twice its health and half its value at range: spend the blade.
Know what the clock is worth. A victory is 0.60. All the damage shaping in the world is 0.30. When theirs.alive is 1, everyone should be on it. When you cannot win, a full-time draw with damage banked still scores and a wipe does not — that is when the last unit alive should kite forever.

Reading the score block

score.team_so_far is the squad's running number, battle_damage_pct and battle_loss_pct are this battle's two shaping terms, and the three weights are handed to you so a prompt can do the arithmetic itself rather than guess at it. armies.ours.hp_pct minus armies.theirs.hp_pct is the cheapest single measure of whether the trade is going your way.

Prompt shape that works

Give the model a rule it can evaluate, not an aspiration. "Take the enemy with the lowest hp among those in range of anybody" is executable; "focus fire" is not.
Branch on the role explicitly. The same prompt runs on a ranger and on a blade, and the two want opposite things at 200 px.
Name the exception. "Unless a melee enemy is within 200 pixels of you, in which case kite" is the difference between a ranger that trades and a ranger that dies.
Say what to do when the battle is decided, in both directions: closing out a win and salvaging a loss are different orders and both are worth points.
Keep it under ~250 words. The board is already in the message; the prompt's job is the policy, not the situation.

What will get you fallbacks

The reply must be a single JSON object beginning with {, with exactly one cogs entry — your own unit. Prose before the object is tolerated, fences are tolerated, "E3" where an integer was asked for is tolerated, and an unknown intent is repaired to focus. What is NOT recoverable is a reply with no JSON object in it at all: that costs a retry, and a second failure costs the turn. The seat plays the published focusfire baseline for that turn, which is not a disaster — but it is not your policy playing.

Check results.fallbackTurns after an episode. A champion with more fallbacks than LLM turns is a prompt bug, not a platform outage.

Testing a prompt

coworld upload-policy coworld-smac-starcraft-micro \
  --name my-smac --run /bin/smac-starcraft-micro-player \
  --secret-env PLAYER_PROMPT="<your strategy>"

Then watch the replay: the match feed prints your note every turn and the focus ring on the board shows, as a shape, whether your squad is actually concentrating fire.