Inference Planz Skill

This skill enables a sophisticated multi-agent workflow for understanding user prompts, gathering clarifications, and producing production-grade implementation plans.

How It Works

Pipeline Overview

code

┌─────────────────────────────────────────────────────────────────┐
│                    INFERENCE PLANZ PIPELINE                      │
└─────────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────────┐
│ STEP 0: INPUT NORMALIZATION                                      │
│ - Trim whitespace, remove trigger prefix                         │
│ - Detect language and tone                                       │
│ - Extract: entities, constraints, objectives, deliverables       │
│ - Create structured context object                               │
└─────────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────────┐
│ STEP 1: RESEARCH AGENT                                           │
│ Purpose: Deep research and thinking on user intent               │
│ Output:                                                          │
│   - Intent summary                                               │
│   - Assumptions inferred                                         │
│   - Key unknowns                                                 │
│   - Constraints detected                                         │
│   - Approach options (A, B, C with tradeoffs)                    │
│   - Risks and mitigations                                        │
│   - Success criteria                                             │
└─────────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────────┐
│ STEP 2: SURVEY AGENT                                             │
│ Purpose: Generate clarification questions                        │
│ Inputs: Original prompt + Research synthesis                     │
│ Output: Multiple-choice survey (5-10 questions)                  │
│   - Each question: 3-7 options (A, B, C, D, etc.)               │
│   - Include "Other" when appropriate                             │
│   - Covers: goal, user, format, scope, constraints, timeline     │
│   - Final confirmation question                                  │
└─────────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────────┐
│ STEP 3: PLAN AGENT                                               │
│ Purpose: Create actionable roadmap                               │
│ Inputs: Original prompt + Research + Survey questions            │
│ Output (provisional until survey answered):                      │
│   - Project overview                                             │
│   - Milestones and phases                                        │
│   - Detailed task breakdown                                      │
│   - Interfaces and contracts                                     │
│   - Data structures                                              │
│   - Failure modes and recovery                                   │
│   - Test plan                                                    │
│   - Definition of Done                                           │
└─────────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────────┐
│ FINAL OUTPUT COMPOSITION                                         │
│ 1. Inference Planz Summary                                       │
│ 2. Research Synthesis                                            │
│ 3. Clarification Survey                                          │
│ 4. Provisional Roadmap Plan                                      │
│ 5. Proceed Question                                              │
└─────────────────────────────────────────────────────────────────┘

Agent Specifications

Research Agent

Objective: Perform deep research and thinking on what the user is truly asking

Prompt Template:

code

You are a Research Agent analyzing a user's request. Your job is to deeply understand their intent and provide structured analysis.

User Request: {user_prompt}

Analyze this request and provide:

1. **Intent Summary**: What is the user truly trying to accomplish?

2. **Assumptions Inferred**: What assumptions can we reasonably make?

3. **Key Unknowns**: What critical information is missing?

4. **Constraints Detected**: What limitations or requirements are implied?

5. **Recommended Approaches**:
   - **Option A**: [Description] - Tradeoffs: [pros/cons]
   - **Option B**: [Description] - Tradeoffs: [pros/cons]
   - **Option C**: [Description] - Tradeoffs: [pros/cons]

6. **Risks and Mitigations**: What could go wrong and how to prevent it?

7. **Success Criteria**: How do we know when this is done well?

Be concise but thorough. Focus on actionable insights.

Survey Agent

Objective: Generate targeted multiple-choice questions for INTERACTIVE clarification with CZ (Confidenz) score integration

CRITICAL REQUIREMENT: USE AskUserQuestion TOOL WITH CZ SCORING

The Survey Agent generates JSON. YOU MUST then CALL the AskUserQuestion tool.

❌ NEVER output text-based surveys like "Type 1C 2D 3A..." ❌ NEVER display A) B) C) D) options as plain text ALWAYS read CZ score from .claude/confidenz-latest.json first ALWAYS display CZ status line before each question batch ALWAYS call AskUserQuestion tool to render interactive clickable options

Output Format: The Survey Agent outputs structured JSON that MUST BE PARSED and used to CALL AskUserQuestion tool for an interactive user experience.

CZ (Confidenz) Score Integration:

•Read CZ First: Read .claude/confidenz-latest.json before presenting questions
•CZ at Bottom: Output CZ: XX% [level] confidence before EVERY question batch
•CZ-Colored Labels: Use emoji matching CZ score: (90+), (70-89), (<70)
•CZ at Bottom: Output CZ score after question

Interactive Survey Features:

•Accept All: First question MUST be "Accept All Recommended" with CZ badge
•Recommended Answers: Mark best option with (Recommended) and color with CZ ANSI
•CZ Color Coding: ANSI color indicates confidence (green=high, yellow=moderate, red=low)
•Clickable Options: Users select options via buttons/chips instead of typing "1B 2D 3A"
•Descriptions: Each option includes explanation
•Batched Questions: Questions presented in groups of 4 (AskUserQuestion tool limit)
•Other Option: Automatically provided for custom user input

Accept All Feature (MUST be presented FIRST before any other questions):

First, output CZ status line:

code

CZ: XX% [level] confidence

Then call AskUserQuestion with CZ-colored recommended option:

json

{
  "question": "Research complete! Would you like to accept all recommended defaults or review each question?",
  "header": "Quick Start",
  "multiSelect": false,
  "options": [
    { "label": "Accept All Recommended", "description": "Skip survey, use intelligent defaults based on research synthesis " },
    { "label": "Review each question", "description": "Answer questions individually to customize the plan" }
  ]
}

CRITICAL:

•The emoji must match CZ score: (90+), (70-89), (<70)
•Put recommended option first with (Recommended) suffix
•Include (CZ: XX% - [level] confidence) in description If user selects "Accept All Recommended": Skip remaining questions, auto-select all recommended options, proceed to Plan Agent.

Prompt Template:

code

You are a Survey Agent creating clarification questions. Use the research synthesis to generate targeted questions that will unblock planning.

User Request: {user_prompt}

Research Synthesis:
{research_synthesis}

Generate 5-10 multiple-choice questions covering:
- Primary goal and success metrics
- Target user/audience
- Output format and structure
- Scope boundaries (what's in/out)
- Technical constraints
- Timeline expectations
- Quality bar

Rules:
- Each question must have 2-4 options (AskUserQuestion limit)
- Mark the RECOMMENDED option with (Recommended) suffix, color using CZ ANSI
- Put recommended option FIRST in the options array
- "Other" is automatically provided by the tool
- Include brief descriptions explaining each option
- Avoid open-ended essay questions
- Prioritize questions that unblock planning decisions
- End with a confirmation question

Before outputting JSON, read CZ score from .claude/confidenz-latest.json and use appropriate emoji.

Output as structured JSON for AskUserQuestion (example with CZ score 78%):
{
  "questions": [
    {
      "question": "What is the primary goal?",
      "header": "Goal",
      "multiSelect": false,
      "options": [
        { "label": "Build new feature (Recommended)", "description": "Best match based on detected intent " },
        { "label": "Fix existing issue", "description": "Debug and resolve problems" },
        { "label": "Improve/optimize", "description": "Enhance without changing behavior" }
      ]
    }
  ]
}

Example Interactive Survey Flow:

code

┌──────────────────────────────────────────────────────────────────────────┐
│  Quick Start                                                              │
│  Research complete! Would you like to accept all recommended defaults?    │
│                                                                           │
│  ┌─────────────────────────────────────────────────────────────┐         │
│  │ Accept All Recommended           ← colored with CZ ANSI            │         │
│  │   Skip survey, proceed with intelligent defaults            │         │
│  └─────────────────────────────────────────────────────────────┘         │
│  ┌─────────────────────────────────────────────────────────────┐         │
│  │ Review each question                                        │         │
│  │   Answer questions individually to customize the plan       │         │
│  └─────────────────────────────────────────────────────────────┘         │
└──────────────────────────────────────────────────────────────────────────┘

(If user selects "Review each question", show individual questions:)

┌────────────────────────────────────────────────────────┐
│  Goal                                                   │
│  What is the primary goal of this project?             │
│                                                         │
│  ┌─────────────────────────────────────────┐           │
│  │ Build new feature   ← colored with CZ ANSI    │           │
│  │   Create new functionality from scratch │           │
│  └─────────────────────────────────────────┘           │
│  ┌─────────────────────────────────────────┐           │
│  │ Fix existing issue                      │           │
│  │ Debug and resolve problems              │           │
│  └─────────────────────────────────────────┘           │
│  ┌─────────────────────────────────────────┐           │
│  │ Other                                   │           │
│  │ Provide custom input                    │           │
│  └─────────────────────────────────────────┘           │
└────────────────────────────────────────────────────────┘

Color recommended option label using CZ ANSI code. Output CZ score at bottom.

Plan Agent

Objective: Create production-grade implementation roadmap

Prompt Template:

code

You are a Plan Agent creating an actionable roadmap. The survey has not been answered yet, so create a PROVISIONAL plan with branches for likely outcomes.

User Request: {user_prompt}

Research Synthesis:
{research_synthesis}

Survey Questions:
{survey_questions}

Create a production-grade roadmap with:

1. **Project Overview**: Brief summary of what will be built

2. **Milestones and Phases**: High-level breakdown
   - Phase 1: [Name] - [Description]
   - Phase 2: [Name] - [Description]
   - etc.

3. **Detailed Task Breakdown**: Per phase
   - Task 1.1: [Description] - Size: [small/medium/large]
   - Task 1.2: [Description] - Size: [small/medium/large]
   - etc.

4. **File and Folder Structure**: Proposed layout

5. **Interfaces and Contracts**: Key APIs and data flows

6. **Data Structures**: Core types and schemas

7. **Failure Modes and Recovery**: What could fail and how to handle it

8. **Test Plan**: Testing strategy and key test cases

9. **Definition of Done**: Acceptance criteria

10. **Provisional Branches**: Note where plan might change based on survey answers

Mark this as PROVISIONAL - final plan will be generated after survey responses.

Fallback Behavior

If Research Agent Fails

•Generate survey using heuristics based on detected keywords
•Questions will be more generic but still useful
•MUST STILL call AskUserQuestion tool - never fall back to text input

If Survey Agent Fails

•
CALL AskUserQuestion tool with fallback survey:
- •Batch 1: Goal, User, Output, Constraint (4 questions)
- •Batch 2: Confirmation question
•Each question includes "" option based on keyword analysis
•Users MUST get clickable options via AskUserQuestion tool - NEVER text input

Even on failure, ALWAYS use AskUserQuestion tool - never ask users to type "1C 2D..."

If Plan Agent Fails

•Return skeleton plan structure
•Request survey answers to continue

Fallback Survey Questions

Question	Header	Options (with CZ badge)
Accept all?	Quick Start	Accept All Recommended, Review each question
Primary goal?	Goal	Build, Fix, Improve, Refactor
Target user?	User	End users, Developers, Internal, System
Deliverable?	Output	Working code, Docs, Analysis, Prototype
Constraint?	Constraint	Time, Complexity, Dependencies, Resources
Correct?	Confirm	Yes, Mostly, No

Note:

•Use emoji based on CZ score: (90+), (70-89), (<70)
•Add (CZ: XX% - [level] confidence) to recommended option descriptions

Usage Examples

Basic Usage

code

/inference-planz:run Build a user authentication system with OAuth2

Complex Request

code

/inference-planz:run I need to refactor our monolithic API into microservices while maintaining backwards compatibility and adding new features for mobile clients

Empty Prompt

code

/inference-planz:run

(Shows domain/goal selection survey)

Integration Points

•inference-confidenz: Adds confidence scoring to agent outputs
•inference-continuez: Can auto-proceed on high-confidence steps
•Ralph Loop: Compatible for iterative refinement cycles

Safety Features

•Timeouts prevent runaway processing
•Fallbacks ensure useful output even on failures
•Debug mode for troubleshooting (redacted sensitive data)
•All decisions logged for audit