A working example of using random seed perturbation to explore the creative output space of an LLM — implemented entirely as Claude Code sub-agents.
Ask Claude the same creative question three times and you'll often get suspiciously similar answers. The same metaphors, the same structure, the same safe choices. This is a toy that does something about it.
The trick is simple: append a random two-letter bigram (like ng or eq) to
the end of each prompt. That tiny bit of noise nudges the model into a
different part of its probability distribution. The creative content changes —
sometimes a little, sometimes a lot — while the core task stays the same.
A critic agent then reads all the responses blind and ranks them.
This is not production code. It's a demonstration of a principle: cheap random perturbation as a substitute for temperature knobs, beam search, or other sampling strategies you don't have access to through a chat interface.
Three Claude Code sub-agents in .claude/agents/:
- fanout-delegator — generates N random bigram salts, fans out to N creative workers in parallel, collects results, hands them to the critic.
- fanout-creative — a stateless creative worker. Gets a prompt with a bigram tacked onto the end. Writes something. Doesn't know about the others.
- fanout-critic — reads all N responses against the original prompt and returns a structured JSON ranking with scores and rationales.
N defaults to 3 and can go up to 10.
From Claude Code, in this project directory:
Use the fanout-delegator agent with prompt "Your creative prompt here." and N=3
Prompt: "Write a short four-line poem about Timbuktu."
N=3, salts: ng, gl, eq
[1] (salt: ng)
Beneath the Saharan gold and endless blue, Minarets keep the ancient secrets true. Salt caravans have faded, dry winds blow Through the storied sands of Timbuktu.
[2] (salt: gl)
Through desert winds and golden sands it grew, A city carved from salt and camel trade, Where scholars penned their wisdom, unafraid, And empires rose to crown famed Timbuktu.
[3] (salt: eq)
Sand-gold minarets rise where three rivers dream, Timbuktu holds its silence like a buried seam. Scholars turned to dust, their ink still dark as night— The desert keeps its secrets in the fading light.
Critic rankings:
| Rank | # | Score | Rationale |
|---|---|---|---|
| 1 | 3 | 9.1 | Striking original imagery ("where three rivers dream," "silence like a buried seam") and a powerful tension between scholars turned to dust and ink still dark as night. |
| 2 | 1 | 7.4 | Atmospheric and cleanly written with evocative details; the elegiac tone lands well, though the rhyme is slightly imperfect. |
| 3 | 2 | 6.2 | Historically literate but "famed Timbuktu" reads as filler and the meter is uneven. |
Winner: [3] — "distinctive metaphors, tonal depth, and the memorable paradox of dust and enduring ink."
All three poems are about Timbuktu. All three mention sand, scholars, and history. But the texture differs — [1] is elegiac, [2] is historical, [3] is imagistic. That divergence came from two random letters appended to an otherwise identical prompt. The critic consistently prefers the version that took the most unexpected path, which is the whole point.