A skill that produces one output at a decision point reduces the human to a gatekeeper. The human's role is approve or reject. There is no real choice because there is only one option. Presenting 3-5 structured alternatives changes the dynamic: the human becomes a curator selecting from options that each represent a different valid interpretation of the brief. That is co-creation, and it produces better final output than approval of a single attempt. This anti-pattern appears across Claude Code skills built on AEM's workflow framework.

TL;DR: Single-output decision points waste the human's attention on threshold assessment rather than judgment. Output variation — 3-5 distinct alternatives at key decision moments — is the highest-impact skill design change available without modifying the core task logic. Selection from options produces better final output because choosing forces clearer criteria than approving one draft.

What is a human decision point and why does it matter?

A human decision point is any step where the skill pauses and asks a human to make a choice before the workflow continues. In a content publishing skill: "Review this draft before I schedule it." In a code review skill: "Which of these proposed changes should I apply?"

Decision points exist because some operations benefit from human judgment — aesthetic choices, strategic tradeoffs, risk assessments that require context the skill does not have. The decision point is where the human adds their value to the workflow.

The design question is: what should the human be evaluating?

With a single output, the human evaluates whether the output is good enough to proceed. That is a yes/no function. It requires no judgment, only threshold assessment. With 3 options, the human evaluates which option best fits their intent, their audience, their constraints. That requires actual judgment — and produces a better downstream result.

"When you give a model an explicit output format with examples, consistency goes from ~60% to over 95% in our benchmarks." — Addy Osmani, Engineering Director, Google Chrome (2024)

Applied to output variation: when you specify the format and diversity criteria for multiple options explicitly in the skill step, the options Claude generates are more consistently distinct and useful.

Why does a single output reduce the human's role?

Single-output decision points underperform because they create approval bias, push Claude toward conservative center-of-distribution outputs, and destroy the interpretive signal that comes from selection. Each of these mechanisms is independent; together they compound into a materially worse final output compared to a choose-from-N design.

Tversky and Kahneman (1974) documented that arbitrary anchors shift estimates by 20 percentage points or more. A single output functions as exactly that kind of anchor: it fixes the human's reference point before evaluation begins.

  • Approval bias: When a human sees one output, they calibrate their judgment against that single output. If it is 80% good, the remaining 20% starts to seem like a reasonable compromise rather than a genuine gap. With 3 options, each one resets the calibration — the human selects the best across alternatives, not the acceptable from a single attempt.

  • Conservative outputs: Claude, when generating a single output, optimizes toward "probably correct" rather than "distinctively right." A single output hedges toward the center of the probability distribution. Multiple options allow Claude to explore the full space of valid interpretations, including bolder choices that a single-output approach would never risk producing.

  • Lost signal from rejection: When a human rejects a single output, the next version is Claude's guess at what the human wanted. When a human selects from 3 options, their selection tells Claude which interpretation axis they preferred. The next iteration starts from a precisely calibrated baseline instead of a guess.

How does output variation work in practice?

The implementation lives in the process steps. Instead of "Generate a draft summary," write: "Generate 3 summary options. Option A: formal tone, under 100 words. Option B: conversational tone, under 150 words. Option C: bullet-point format, 5 bullets maximum. Present them numbered. After the human selects one, proceed to the next step."

The criteria for each option should differ along a meaningful axis: tone, length, format, framing, or emphasis. Three options that are 95% identical are not real options — they are the same output with minor variations. Meaningful variation requires that each option makes a different tradeoff explicitly.

In our production skills, we specify the variation axis in the step instruction, not just the count. "Generate 3 alternatives with different core claims" produces more useful variation than "Generate 3 alternatives." The axis is the design decision. Boris Cherny, who built Claude Code at Anthropic, reports that AI agents which verify their own work improve output quality by 2-3x compared to those that do not (Cherny, Anthropic, 2024). The same logic applies to output variation: giving the selection step a clear verification criterion produces measurably more distinct alternatives than leaving the axis implicit.

For how to structure multi-phase workflows where each phase ends at a decision point, see How Do I Write Step-by-Step Instructions for a Claude Code Skill?.

What are the three types of decision point — and which needs output variation?

Skill design methodology identifies three checkpoint types: approve/decline gates for irreversible actions, choose-from-N for creative and strategic decisions, and open-field for cases where the human intends substantial modification. Output variation belongs exclusively at the choose-from-N checkpoint; presenting 3-5 alternatives is correct there, not at gates or open-field steps.

  • Approve/decline: Single output, yes/no decision. This is appropriate only for irreversible actions where proceeding without human sign-off would be unacceptable: sending an email, publishing to production, executing a database write. Use it for gates, not for creative or strategic decisions.

  • Choose-from-N: 3-5 options presented for selection. This is the standard decision point for content, strategy, and aesthetic choices. The human selects one; the workflow continues from the selected option. This is where output variation belongs.

  • Open-field: The skill presents a draft and invites the human to modify it. Appropriate when the human's contribution is likely to be substantial rather than selective — when they know exactly what they want to change. Use sparingly.

The mistake is applying approve/decline logic to creative decision points. A content publishing skill that stops at "Approve this headline?" has misapplied the gate pattern to a choice that would produce a better result with 3 alternatives. Research on prompt format sensitivity found that GPT-3.5-turbo performance varies by up to 40% depending on how output instructions are structured (He et al., arXiv 2411.10541, 2024). This is consistent with the argument that how you frame the options step matters as much as whether you include it.

For a broader look at advanced patterns in production skills, including the verifier pattern, see What Advanced Skill Design Patterns Exist Beyond Basic SKILL.md Files?.

FAQ

How many options is optimal at a decision point?

3-5. Fewer than 3 limits the choice space. More than 5 creates decision fatigue — humans presented with 7+ options tend to default to the middle option rather than engaging with the full range. Cowan (2001, Behavioral and Brain Sciences) established that working memory holds 3-5 meaningful chunks; option sets beyond that range exceed the capacity for genuine comparative evaluation. Iyengar and Lepper (2000, Journal of Personality and Social Psychology) found that shoppers offered 6 options purchased at a 40% rate, versus 3% when offered 24. That is a 13x drop in conversion from excess choice.

Can a skill have multiple decision points with options at each?

Yes. Multi-phase workflows benefit from output variation at each phase transition: options at the outline stage, then options at the draft stage. The human who selected outline Option B now selects from 3 draft variations that interpret that outline differently. Each selection sharpens the final output incrementally.

Is output variation always better than a single output?

No. For deterministic tasks — date formatting, code compilation, data transformation — there is one correct output and presenting options creates confusion. Output variation is a design pattern for tasks where multiple valid outputs exist and human judgment adds value to the selection process.

Does requesting 3 options increase token cost?

Yes, generating 3 options costs approximately 3x the tokens of generating 1 option at that step. For skills running at high frequency, factor this into the economic model. The tradeoff is whether better output quality justifies the additional token cost. For most creative and strategic tasks, it does.

How do I tell Claude to present options in a consistent format?

Specify the format in the process step: "Present each option as a numbered header followed by the option content. After all three options, ask the user to type a number to select." Explicit format instructions in the step text produce more consistent option presentation than leaving format to Claude's discretion.

What if my client only wants one output — is output variation still the right design?

That preference is worth testing. Clients who prefer single outputs often haven't experienced well-designed output variation — they're used to the choose-from-3 pattern producing near-identical options. If the options are genuinely distinct and the axes of variation are meaningful, most clients revise their preference after the first session.

Last updated: 2026-04-19