# Skill: Identify AI model families and choose the right one for a given task

> Lets an AI coding/design agent quickly orient in the Claude / GPT / Gemini / open-model landscape and recommend or switch to the correct family based on task type, privacy needs, and budget; applies whenever a model selection decision needs to be made.

Source: sodigi·learn — models/semeystva-modeley · https://sodigi.io/learn/models/semeystva-modeley

## When to use
- A user asks "which AI model should I use for X?"
- Evaluating a new model announcement to decide whether it changes current tooling.
- Designing a workflow that mixes multiple models across tasks.
- Explaining why one model outperforms another on a specific job.

## Core rules
- There are two camps: closed (API/web only -- Claude, GPT, Gemini) and open (downloadable weights -- Llama, Qwen, Mistral, DeepSeek). Understand the camp before picking the model.
- "Open weights" means you hold the file and can run it anywhere; it does not mean "do whatever you want" -- check the license.
- Memorize family temperament, not version numbers; versions go stale in weeks.
- There is no single best model; the market is split into niches -- pick by task.
- Claude: precise text, long documents, code quality, calm tone; best fit for writing and code review.
- GPT: breadth and ecosystem; most tutorials and integrations exist here; solid all-rounder.
- Gemini: multimodality (text + image + video) and Google-services integration.
- Open models (Llama, Qwen, Mistral, DeepSeek): free to run locally, full privacy control; require hardware and setup effort.

## Procedure
1. Identify the dominant requirement: precise writing/code -> Claude; breadth/ecosystem/tutorials -> GPT; images/video/Google services -> Gemini; privacy/on-premises/zero cost -> open model.
2. Identify the access constraint: no install, just a browser -> any closed model; sensitive data that cannot leave premises -> open model locally.
3. Identify the budget constraint: zero budget, own GPU available -> open; subscription or pay-per-token fine -> closed.
4. If undecided between two candidates, run the exact same task through both and compare answers blind (hide the model name while judging).
5. Resist updating the recommendation every time a new version ships; the family temperament changes slowly.

## Ready-to-use prompt
```
Act as an experienced mentor. Help me with the task below.

Task: [describe your task in one or two sentences]
What matters: [for example - precision, speed, long text, analyzing an image, privacy]

Do the work, and at the end briefly explain what you based it on.
```

## Pitfalls
- Chasing "number one in the world" headlines -- benchmark rankings are niche-specific and change weekly.
- Memorizing version numbers instead of family temperament -- 3.7 vs 3.5 matters less than "is this a Claude or a GPT job?"
- Conflating "open" with "free in every way" -- you pay for hardware, electricity, and sometimes licensing restrictions.
- Assuming a beginner can easily self-host an open model -- without a capable GPU and some technical know-how it is a significant hurdle.
- Trusting a single benchmark chart -- it is averaged over thousands of other people's tasks, not yours.
- Using one model for everything when a different tool would be clearly stronger for a sub-task.