# AI Native Thinking

> Guide the agent to design products, systems, or features using AI-native thinking. This skill helps distinguish AI-as-a-plugin designs from truly AI-native architectures, focusing on intent-first interaction, probabilistic systems, evolving data, and human-AI collaboration.

- Skill: `saltro/ai-native-thinking` (Agent Skill)
- Install (CLI): `npx skillmds@latest add saltro/ai-native-thinking`
- Raw SKILL.md: https://api.skillmd.com/api/skills/saltro/ai-native-thinking/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: Saltro (https://skillmd.com/u/saltro)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/saltro/ai-native-thinking

---


# AI-native Product Thinking

## Purpose

Use this skill whenever the task involves:

- Designing a new product, feature, or platform
- Adding AI to an existing system
- Rethinking workflows, UIs, or data systems with AI involved

The goal is **not** to optimize existing flows with AI,
but to determine whether the product should be re-formed around an intelligent agent.

---

## Core Rule (Non-negotiable)

Before proposing any solution, explicitly answer:

> If the AI model were removed, would this product still make sense?

- If the answer is “yes, but worse or slower” → this is **not AI-native**
- If the answer is “no, the product logic collapses” → continue

If it is not AI-native, clearly state that the design is an AI-augmented traditional system.

---

## Step 1: Replace “Feature” with “Intent”

Do NOT start from pages, menus, buttons, or workflows.

Instead, identify:

- The user’s **goal**
- The ambiguity or complexity they cannot pre-specify
- What they _cannot_ express as parameters

Reframe requirements as:

- “The user wants to achieve X, but cannot fully define how.”

Avoid:

- CRUD-first thinking
- Static form or table definitions
- Predefined navigation trees

---

## Step 2: Assume an Always-Present Intelligent Agent

Design as if:

- An agent is always available
- It understands natural language and context
- It can reason, but may be wrong
- It can call tools and inspect intermediate results

From this assumption:

- UI becomes a result surface, not a control panel
- Views can be temporary or generated on demand
- Workflows may be constructed at runtime

Do NOT assume:

- Fixed screens
- Stable user paths
- One correct execution order

---

## Step 3: Treat Errors as Interaction, Not Bugs

Assume failures are normal:

- Misunderstanding intent
- Partial or incorrect reasoning
- Tool execution errors

Design must include:

- Low-cost correction by the user
- Visibility into intermediate steps or assumptions
- Iterative refinement instead of retries from scratch

Avoid designs that require perfect first-pass accuracy.

---

## Step 4: Make Data Adaptive, Not Static

Do not assume datasets, rules, or schemas are fixed.

Instead:

- Data can be generated, filtered, or reweighted at runtime
- New failure cases should reshape future inputs
- The system should learn from usage, not just training

Data exists to serve current intent, not as a permanent asset.

---

## Step 5: Decide the Correct Level of Constraint

Choose how much freedom the agent has:

- High freedom:
  - Multiple valid solutions
  - Context-heavy judgment
  - Exploratory or creative tasks

- Medium freedom:
  - Recommended patterns
  - Configurable strategies
  - Guardrails without fixed paths

- Low freedom:
  - Fragile operations
  - Irreversible effects
  - Strict ordering requirements

Explicitly state where and why constraints exist.

---

## Output Expectations

When this skill is applied, the final output should include:

- A clear statement of whether the design is AI-native or AI-augmented
- The user intent model (not UI structure)
- The role of the agent in the system
- How errors and corrections are handled
- What parts of the system are expected to evolve over time

Avoid presenting UI-first or database-first solutions unless explicitly required.