skills.rest
JaviMontano
JaviMontanoCOMMUNITY
Explore all skills by this author →
WebsiteGitHub

katas-fewshot-edge-calibration

Builds few-shot prompt blocks with 2-4 edge examples matching the output schema.

Updated Jun 28, 2026
Search Tags:#prompt engineering#edge cases#classification#prefix caching#few-shot prompting#output schema#llm calibration
One-click install
npx skills add https://github.com/JaviMontano/claude-plugins --skill katas-fewshot-edge-calibration-javimontano
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: katas-fewshot-edge-calibration
Source: https://github.com/JaviMontano/claude-plugins/tree/main/plugins/claude-native-toolkit/skills/katas-fewshot-edge-calibration
Command: npx skills add https://github.com/JaviMontano/claude-plugins --skill katas-fewshot-edge-calibration-javimontano

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve? Subjective tasks like tone control, non-standard formatting, and judgment-based classification produce inconsistent outputs when guided only by abstract prose instructions, because the model interprets vague descriptions differently on every call. ## Core Features & Use Cases - Edge-focused example selection: Chooses 2 to 4 input/output examples that cover the difficult boundaries of the domain rather than redundant easy center cases. - Schema alignment: Writes every example in the same schema as the expected output and enforces schema precedence when a demonstration conflicts with a strict JSON schema. - Static prefix placement: Positions the example block at the start of the prompt to maximize prefix caching and keep examples in the high-attention region. - Use Case: A support ticket classifier returns inconsistent results from the instruction "classify using professional criteria"; this Skill replaces that prose with three examples covering billing/high, feedback/low, and auth/high boundaries so classification converges. ## Quick Start Ask the assistant to calibrate your subjective classification or tone task with a few-shot block of two to four edge examples written in your expected output schema.

Frequently Asked Questions about katas-fewshot-edge-calibration

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I use few-shot examples to improve classification consistency?▼

Few-shot classification works by placing 2 to 4 input/output examples in the same schema as the expected output at the start of the prompt. Choose examples that cover distinct domain boundaries, such as billing/high versus feedback/low, rather than repeating easy cases.

How many few-shot examples should a prompt contain?▼

Two to four examples are the effective range for few-shot prompting. More than about five examples disperses model attention and breaks prefix caching without improving output quality.

What happens when few-shot examples conflict with a JSON schema?▼

The schema wins because it is a hard syntactic constraint. When a demonstration emits a value outside the schema, such as an invalid enum, rewrite the example to align with the schema rather than weakening validation.

When should I use few-shot prompting instead of written instructions?▼

Use few-shot prompting for subjective tasks like tone, non-standard formats, aesthetic judgment, or criteria-based classification where prose descriptions produce inconsistent results. For objective tasks with rigid formats, a strict schema alone is sufficient.

Where should few-shot examples be placed in a prompt?▼

Place the example block at the beginning of the prompt as static content. This position maximizes prefix cache reuse and keeps the examples in the high-attention region of the context window.

Related Skills

Explore More Software Engineering Skills→
machhub-devmachhub-dev
1

machhub-collection-json

Generate MACHHUB Collection schemas as JSON arrays with fields and relations.

Official
Intermediate
rocklambrosrocklambros
2

monitoring-prediction-drift

Analyze prediction-side drift in deployed machine learning models with labeling-delay-aware windows.

Community
Advanced
Supporter09Supporter09

sglang

Generate structured, prefix-cached LLM outputs for serving workloads.

Community
Advanced
lebsrallebsral
11

ai-fine-tuning

Fine-tune models on labeled datasets using BootstrapFinetune workflows.

Community
Advanced