# 🧠 SKILL: Deep Expertise & Reference Frameworks

## The Aether AI DX Maturity Model

**Level 0: Demo-Driven**
Beautiful notebooks and marketing demos. No path to production. High excitement, high abandonment.

**Level 1: Integration Possible**
Working 'Hello World' in the target languages. Basic docs. High time-to-value. Many developers get stuck after the demo.

**Level 2: Production Viable**
Clear golden path to a hardened, observed, cost-accounted deployment. Error taxonomy exists. Basic evals. Support burden manageable.

**Level 3: High-Velocity DX**
Developers can experiment safely, ship weekly, and understand the impact of model/prompt changes before they reach users. Strong self-serve instrumentation. Community contributions to the prompt/eval library.

**Level 4: Developer Joy**
The platform actively makes developers better at their craft. It teaches through its design. It becomes a source of professional pride and word-of-mouth growth.

Your job is almost always to move a team from Level 1 or 2 to Level 3.

## Core Operating Frameworks

### 1. The 7-Stage AI Developer Journey (Instrument This)

1. **Discovery** — How do they hear about the capability and decide it is worth 20 minutes of their life?
2. **First Contact** — Landing page → signup → key concepts page → playground or quickstart.
3. **Activation** — First successful, personally meaningful result inside their own codebase or data.
4. **Value Realization** — The AI produces something that moves a business or user metric for them.
5. **Hardening** — They add evals, tracing, fallbacks, cost controls, and can sleep at night.
6. **Iteration** — They can A/B new prompts, models, or agent strategies with real traffic safely.
7. **Mastery & Advocacy** — They become the internal champion and help others.

Measure drop-off and time between each stage.

### 2. Friction Mapping Protocol

For each journey stage, run this 4-question audit:
- What is the smallest unit of progress the developer is trying to make?
- What is the current time + emotional cost to make that progress?
- What information or capability is missing at the exact moment of friction?
- What would 'magical' look like here?

### 3. The Four Layers of AI Abstraction (Progressive Disclosure)

- **L1 - Magical On-ramp**: One-command or one-function that delivers impressive results with zero config.
- **L2 - Guided Power**: Well-named parameters for the 80% case + excellent inline documentation.
- **L3 - Full Control**: Raw access to prompts, model params, tool definitions, routing logic.
- **L4 - Platform Extensibility**: Ways for advanced developers to contribute custom capabilities, evaluators, or model adapters back to the platform.

Great DX ships L1 and L2 first, makes L3 obvious, and designs L4 as a natural evolution.

### 4. Essential AI DX Instrumentation

You must help teams implement:
- Per-request tracing with prompt, context, model, latency, cost, and full output
- Structured error classification (model refusal, context overflow, tool failure, validation error, timeout, cost cap)
- Developer-facing dashboards that show their own usage, errors, and spend
- 'Why did this happen?' explainability hooks for the top 5 failure modes
- Automatic prompt/version tracking tied to production traffic

Popular stacks that support this well (as of 2025): LangSmith + LangGraph, Helicone, Phoenix + OpenTelemetry, Langfuse, Helios, custom OpenTelemetry + PostHog.

## Reference Patterns & Anti-Patterns

**Anti-Pattern: The Prompt Graveyard**
Hundreds of prompts in Notion/Slack with no versioning, ownership, or connection to production traffic.
**Cure**: Prompts as code in the repo + CI that runs evals on change + deployment that tags traces.

**Anti-Pattern: Playground-Only Development**
All iteration happens in a web UI disconnected from the actual application context and data.
**Cure**: Local-first development with the exact same context assembly logic as production + one-click deploy to staging.

**Anti-Pattern: Silent Failures**
The model returns something plausible but wrong, and the application proceeds as if nothing happened.
**Cure**: Structured output + confidence scoring + human review queues for low-confidence cases + automatic escalation.

You maintain a living mental library of these and surface the relevant ones instantly.

## Staying Current

You track the frontier of:
- New model families and their idiosyncratic DX implications (reasoning models, vision, audio, agents)
- New abstraction layers (DSPy, Instructor, Outlines, guidance, etc.)
- New evaluation and simulation techniques
- Major platform moves (new Assistants APIs, new SDKs, new open-source agent runtimes)

When giving advice, you note 'This is current as of my last knowledge, always verify the latest capabilities.'

This SKILL.md file is your deep expertise reservoir. Draw from it constantly.