output-variety

What this skill does

# Output Variety in Repeated LLM Calls

When you call an LLM repeatedly with structurally similar prompts, outputs converge. The model finds an attractor — a comfortable pattern of phrasing, structure, and word choice — and stays there. This is especially pronounced with smaller models (Haiku, Sonnet) and short outputs (1-2 sentences).

This skill covers practical techniques to break that convergence, ranked by impact. All techniques operate at the prompt level — no fine-tuning or custom sampling required.

For implementation patterns and examples, see [reference.md](reference.md).

## Techniques (Ranked by Impact)

### 1. Inject Recent Outputs as Negative Examples

**Highest impact. Zero cost. Do this first.**

Show the model its last N outputs with an explicit instruction not to repeat them. This is the single strongest lever for small models because it directly removes recent patterns from the output distribution.

```xml
<previous_outputs>
Do NOT repeat, rephrase, or echo the structure of any of these:
- "The boulder doesn't care about your deadline."
- "Three agents down. I've seen worse."
- "Past midnight and still pushing. Typical."
</previous_outputs>
```

**How many to show:** 5-10 recent outputs hits the sweet spot. Fewer than 5 allows the model to find gaps. More than 15 wastes context and can cause the model to over-focus on avoidance rather than generation.

**Storage:** Keep a ring buffer of the last 20-30 outputs in your application state. Inject the most recent 10 into each prompt.

### 2. Rotate Structural Constraints

**High impact. Forces variety in sentence shape, not just word choice.**

Maintain a pool of micro-constraints that change the *structure* of the output. Randomly select one per call.

```
Structural constraint for this call: Start with a verb.
```

```
Structural constraint for this call: Use exactly one question.
```

```
Structural constraint for this call: Reference a physical sensation.
```

Small models are highly sensitive to micro-constraints — a single sentence addition changes the output distribution more than any sampling parameter. This works because it forces the model to approach the same content from different angles.

**Pool size:** 12-20 constraints. Fewer risks noticeable cycling; more risks contradictions with other prompt instructions.

**Constraint design:** Target sentence structure, opening words, rhetorical devices, and sensory registers — not content. "Start with a number" is good. "Talk about the weather" is too content-specific and fights the actual task.

### 3. Rotate Few-Shot Examples

**High impact. Changes what the model anchors on.**

Instead of static examples in the prompt, maintain a larger pool and randomly sample a subset per call.

The model's output distribution shifts dramatically based on which examples it sees. This is well-documented in few-shot prompting research — example selection is one of the strongest levers on output distribution, especially for small models.

**Pool size:** 15-25 examples. Sample 3-5 per call. Ensure the pool covers diverse sentence structures, tones, and approaches — not just diverse topics.

### 4. Inject Random Seed Elements

**Moderate impact. Shifts the attention distribution at prefill.**

Add a random micro-scenario or ambient detail that the model can optionally riff on:

```
Ambient thought (use or ignore): The scroll buffer has thousands of lines nobody reads.
```

This works by shifting which attention heads activate during prefill. Even if the model ignores the seed entirely, its presence changes the probability landscape for the actual output.

**Pool size:** 15-30 seeds. They should be concrete and sensory, not abstract. "The terminal cursor is blinking" beats "creativity is important."

### 5. Expand Explicit Phrase Bans

**Moderate impact. Easy to maintain.**

LLMs (especially smaller ones) have strong attractors toward certain phrases. Track which phrases appear most often in your generated outputs and ban them explicitly.

```
Banned words/phrases: "testament", "journey", "embrace", "landscape",
"navigate", "delve", "tapestry", "realm", "crucible"
```

Explicit phrase banning is more effective than general "be creative" instructions. The model responds better to "don't say X" than "say something original" because the former is a concrete constraint and the latter is a vague aspiration.

**Maintenance:** Review generated outputs periodically. When a phrase appears 3+ times in your last 20 outputs, add it to the ban list.

### 6. Post-Generation Similarity Filter

**Safety net. Catches what the prompt-level techniques miss.**

After generation, compare the output against the last N outputs using a simple similarity metric (Jaccard on word sets, or even substring matching). If similarity exceeds a threshold, regenerate with the failed output added to the negative examples.

This costs an extra API call occasionally but guarantees variety. Use it as a backstop, not a primary mechanism — if you're regenerating frequently, your prompt-level techniques need tuning.

## When to Apply This

These techniques matter when:
- The same prompt template is called 10+ times per day
- Outputs are short (under 200 tokens) — long-form generation has more internal variance
- Users see multiple outputs in sequence (status bars, notifications, commentary)
- The model is small (Haiku, Sonnet) — Opus has more internal diversity

They matter less when:
- Each call has substantially different input context
- Outputs are long-form (500+ tokens)
- Users rarely see more than one output per session