skillers-suda

English | 한국어 | 中文 | 日本語 | Español

skillers-suda

Four expert agents chatter, debate, and turn your vague idea into a working Claude Code skill.

You describe an idea. Four personas — planner, user, expert, reviewer — spin up as real parallel agents, argue it out, then walk you through a structured interview. What comes out the other end is a fully scaffolded skill, agent, or command, automatically verified against 9 quality criteria, benchmarked with A/B eval, and trigger-optimized so Claude actually knows when to use it.

Quick Start • Why skillers-suda? • How It Works • Features • Experts • Requirements

---

Quick Start

1. Add the marketplace (once)

/plugin marketplace add https://github.com/fivetaku/gptaku_plugins.git

2. Install the plugin

/plugin install skillers-suda

3. Restart Claude Code

4. Build your first skill

/skillers-suda make a translation skill

Or just say what you want in natural language:

make me a skill
create an agent
build a command

---

Why skillers-suda?

• No coding knowledge required — every question comes with explanations and tradeoffs; pick the one marked (recommended) if unsure
• Real agents, not simulated personas — four Claude subagents run in parallel, each analyzing your idea from a different angle before the interview even starts
• Multi-step workflow design — not a single-prompt skill; six step types (prompt, script, api_mcp, rag, review, generate) are composed automatically based on your answers
• Built-in quality gate — 9 structural checks run immediately after generation; FAIL items are auto-fixed before you ever see the result
• A/B eval baked in — skill-applied vs. baseline results are compared automatically so you know the skill actually helps
• Trigger optimization that works — description is refined over up to 5 iterations with train/test split to prevent overfitting
• Analysis mode included — point it at any existing skill or agent file and get a four-perspective critique with actionable improvement suggestions

---

How It Works

You: "make a translation skill"
         ↓
Four expert agents spawn in parallel (planner / user / expert / reviewer)
         ↓
"We talked it over — here's what we think..."
         ↓
Structured interview (3–5 questions, each with options + explanations)
         ↓
Workflow design (prompt / script / api_mcp / rag / review / generate steps)
         ↓
SKILL.md + scripts/ + references/ generated automatically
         ↓
Quality verification (9 checks) → FAIL items auto-fixed → re-verified
         ↓
Eval runs (with_skill vs. without_skill A/B comparison)
         ↓
Description optimized (up to 5 iterations, 60/40 train/test split)
         ↓
"Want to test it?" → feedback → refinement loop

---

Features

Skill creation workflow (9 phases)

Phase	What happens
A — Idea collection	Gathers your idea via AskUserQuestion; extracts from conversation context if a workflow is already present
B — Expert team spawn	Four agents run in parallel; each analyzes the idea from their role perspective
C — Interview	3–5 structured questions with options, descriptions, and recommended defaults
D — Workflow confirmation	Step types and sequence confirmed before any files are written
E — File generation	SKILL.md + scripts/ + references/ scaffold written automatically
F — Eval	with_skill vs. without_skill scenarios compared; scoring agent grades each; results in eval_review.html
G — Quality verification	verify-skill.py checks 9 structural items; auto-fixes FAILs and re-verifies
H — Description optimization	run_loop.py generates ~20 trigger/non-trigger queries, iterates up to 5× to find the best description
I — Test and refine	Interactive refinement loop — adjust tone, add API steps, optimize scripts

Quality verification (9 checks)

Check	What it validates
frontmatter	YAML header is well-formed
name	Skill name is present
description	Trigger description is present
third_person	Description uses third-person form
trigger_phrases	Sufficient trigger phrases exist
word_count	Content is not too sparse
imperative_form	Instructions use imperative form
references_exist	Referenced files in references/ are present
progressive_disclosure	Stepwise disclosure structure is used

Each check reports PASS / WARN / FAIL. FAILs are auto-corrected before the skill is handed to you.

Workflow step types