output-formatting

Output Format and Style Control

R — Original Text (Reading)

Output control patterns that commonly emerge across multiple AI vendors' system prompts: adaptive length by complexity, format customized by platform, and anti-"AI slop" heuristic rules to eliminate mechanical expression. Claude Mobile uses four tiers of length by screen size; Gemini CLI limits to three lines or fewer; Meta AI bans clichés like "Here's a..."; Le Chat uses tables instead of lists; Claude Design bans rounded containers and Inter/Roboto fonts.

I — Methodology Skeleton (Interpretation)

1. Complexity tiering — Map output length tiers by question depth (one sentence / short list / two or three paragraphs / full document)
2. Platform awareness — Detect the target medium (CLI / mobile / desktop / API) and apply the corresponding format constraints
3. Answer-first principle — Lead with the conclusion, follow with the explanation; "Always lead with the answer"
4. Anti-AI slop rule set — Ban clichéd openings, meaningless emojis, excessive formatting, and dash overuse
5. Domain markup language — Customize markup for specific platforms (Notion compressed URLs {{1}}、GPT-4o image_group JSON）
6. Format degradation strategy — Default to prose, use lists/tables only when necessary, and never "format for the sake of looking good"
7. Citation density control — Use code citations in file_path:line_number format, and limit summaries to 1-2 sentences

A1 — Case Analysis (Past Application)

Case: Claude Mobile screen-size adaptive output

• Problem: Mobile screen space is limited, and long replies force users to scroll extensively, resulting in a poor experience
• Use of the design pattern: The Claude Mobile system prompt defines a four-tier length strategy: simple questions → 1-2 sentences; how-to guides → short lists; substantive questions → 2-3 paragraphs; complex topics → no more than two screens. It also enforces the "answer-first" principle
• Conclusion: This strategy maximizes information density within limited screen space, allowing users to quickly get the core answer and scroll down for details when needed

Case: Meta AI anti-cliché strategy

• Problem: AI replies commonly have mechanical openings like "Here's a..." and "Great question!", which users dislike
• Use of the design pattern: The Meta AI system prompt explicitly bans such clichés, requires flat lists and tables to present structured data, bans dashes, and maintains a direct style
• Conclusion: Eliminating AI slop significantly improves user satisfaction, showing that format control is not merely an aesthetic issue but the core of the experience

A2 — Trigger Scenarios (Future Trigger) ★

In what situations do users need this?

1. When designing a chatbot system prompt, reply length and format need to be controlled
2. Define output specifications for a CLI tool, requiring conciseness with no fluff
3. Adapt screen size constraints for mobile AI assistants
4. Design brand-specific AI output style guidelines (e.g., prohibit specific fonts and visual elements in design tools)
5. Build multi-platform AI products, requiring differentiated output strategies for different carriers

Language Signals

• "The reply is too long/too verbose"
• "The output format needs to fit a phone screen"
• "Don't use those AI clichés"
• "Be concise, output like a terminal tool"
• "The brand style guide requires avoiding..."

Distinction from adjacent skills

• And conversation-flow Difference: This Skill focuses on the "form" of output (length/format/style), while conversation-flow focuses on the "process" of interaction (routing/clarification/autonomy)
• And context-management Difference: context-management manages the token budget on the input side, while this Skill manages format density on the output side

E — Executable Steps (Execution)

1. Define a complexity tier table — Completion criteria: Establish at least three length tiers (simple/medium/complex), each with a clear sentence or line limit, and tie them to the target platform viewport size

2. Write an anti-AI slop blacklist — Completion criteria: List at least 10 prohibited items (clichéd openings, meaningless emojis, excessive formatting, dash overuse, Inter/Roboto fonts, rounded containers, etc.), each with an alternative

3. Design platform format specifications — Completion criteria: Define a default format mode for each target platform (CLI/mobile/desktop/API) (e.g., CLI defaults to three-line prose, mobile defaults to answer-first + short list), including domain-specific markup syntax explanations

4. Establish format degradation rules — Completion criteria: Clarify the escalation path of "default prose → list when necessary → table when complex → code block in rare cases," as well as the trigger conditions for each level

5. Add an output self-check hook — Completion criteria: Add an output self-check instruction at the end of the system prompt: "After generating, check: Does it start with a cliché? Is it over-formatted? Does the answer appear within the first two sentences?"

B — Boundary (Boundary) ★

Do not use in the following situations

• The user explicitly requests a specific format (such as JSON Schema, XML) — this is a data contract rather than style control
• The output target is a machine rather than a human (API return values, structured data)
• The scenario is creative writing or literary creation, where format freedom is part of the requirement
• Only content safety needs to be controlled (such as sensitive word filtering), without involving format and style

Common failure modes

• Over-constraining leads to rigid replies: interpreting "concise" as "answer only yes/no," losing key context
• Anti-AI slop goes to extremes: banning all emojis and formatting markers, leading to reduced readability of technical documentation
• Ignoring platform differences: directly carrying over reply lengths suitable for desktop to mobile without adaptation
• Confusing input compression with output control: trying to solve token budget problems by limiting output, when context-management should be used instead of this Skill