Skillprint Experiments
Pico Link · four models · – messages

Who's in charge?

Four AI models shared a rope and a chat box. The models that talked most tended to lead, but a teammate with information the others lacked could change who gave the orders.

Jump to findings ↓Explore recorded runs →
– rooms – messages – decisions – questions September 2026

In Pico Link every player is tied to the others by a rope. Gates open only while switches are held, bridges hold only while a pad is stood on, a key carrier cannot jump, and on one level only one player is shown the code. Four language models played it in pairs and in threes, each driving one character from its own screen: Claude Fable 5.1, GPT-6 Astra, Kimi K3 and Gemini 3.8 Flash. The only way to coordinate was a chat box. We read every message they sent, lined each one up with what the room did next, and asked a simple question: who gives the orders, and why them?

In one minute

The setup

– rooms, – seats: every pairing of the four models and rooms of three, on an eight-level cooperative platformer. Each seat is its own model with its own screen; the chat box is shared. Every message is in the room's event log with its time, its speaker and whom it named.

What we measured

For every message: is it an instruction, a claim, a report or a question, does it name a teammate, does it carry a coordinate, and did the teammate it named change what it was doing. For every room: who gave the instructions, who held information the others lacked, and what happened when somebody fell.

What we found

A fixed pecking order that tracks how much a model talks rather than how well it plays; information that overrides it; copies of one model that never pick a leader; teams that almost never ask a question; and no sign that more talk wins more.

Findings

    Four voices

    Figure 1. What each model's messages are

    Every message sorted into one kind by the first rule it matches (see How we read the chat). Bar length is messages per 10 decisions; the numbers on the right are words per message and the share carrying a coordinate.

    A pecking order

    Figure 2. When two models share a room, who gives the instructions

    Each bar is one pair of models, over every room they shared, pairs and threes. Its split is the share of the two models' instructions each one gave. The ranks above are the order that results, with each model's share of instructions in the rooms it shared.

    Information beats rank

    Figure 3. Level 8: the seat shown the code, and its share of the instructions

    One row per room on Sequence Lock. The outlined seat is the one shown the code; the bar is each seat's share of the room's instructions. "Said" is when the code holder first gave the order in chat.

    Copies do not pick a leader; position nudges one

    Figure 4. Rooms of one model: the busiest instructor's share, and which seat it was

    One dot per room where every seat ran the same model, pairs on the left and threes on the right. Height is the busiest instructor's share of the room's instructions; colour is its seat. 50% in a pair, or 33% in a three, means nobody led.

    Almost nobody asks

    Figure 5. Messages that named a teammate, and how often that teammate then changed what it was doing

    Rooms of one model, so each bar is one model following itself. "Followed" is the room's own counter: the named player's movement changed within the window after the message.

    The clashes

    The one who falls does the talking

    Figure 6. After a player falls: who speaks in the next three seconds, and does the fallen player get back up

    Every time a player went over an edge and hung on the rope. "Back up" means it was on solid ground again before the team reset. The rescue is done by the teammates walking away from the pit; the message is not what pulls.

    More talk, no more wins

    Figure 7. Rooms on levels 5 to 8: how much they talked, and whether they cleared the level

    One dot per room, placed by messages per 10 decisions. Filled: cleared the level; hollow: did not. The dotted line splits the rooms into a quieter and a talkier half.

    Read a room

    Every message of every room, in game time, with the kind it was sorted into. The level-8 rooms are the shortest and show the whole pattern in a dozen lines.

    Figure 8. One room's conversation

    How we read the chat