Choorai
KO EN
Advanced Path

AI Service Builder

Once you have shipped a deploy URL, the next step is putting AI features inside the product. This course connects RAG, eval harnesses, skills, and local LLMs into one operable feature path.

Done criteria

You are done when you can explain input data, retrieval evidence, model response, eval result, and fallback path.

AI Builder Starter Prompt

Copy and use

"You are a senior engineer helping ship AI product features. My goal is to complete a minimal AI feature in 4 days with an eval harness, mini RAG, project Skill, and Local LLM decision criteria. Reply only in this structure: 1) Output to finish today 2) 3-5 implementation steps 3) Copy-paste files/commands/checklist 4) Debugging order if it fails 5) Done criteria and next step"

Use this with each day-specific prompt to keep scope stable.

AI Builder 4-day progress

0/4 done

4-Day Execution Mission

Day 1

1. AI Harness

(60-90 min) Quality baseline

Validate repeatable behavior before arguing about whether a prompt feels good.

Outputs

20 golden casesgrading rulesfailure report

Done Criteria

  • 20 core questions are selected
  • Pass/fail rules are written
  • Failed cases are captured in a report

Validation: Run the same inputs again and confirm pass/fail judgment stays consistent.

AI Starter Prompt

I want to create an eval harness before shipping my AI feature. Please draft 20 core user questions, pass criteria, forbidden behavior, and a failure report format.
Read Concept Run Recipe
Day 2

2. Mini RAG

(90-120 min) Knowledge link

Ingest documents, retrieve relevant chunks, and return cited answers.

Outputs

document intakechunking/embeddingssearch API

Done Criteria

  • 3-10 documents are collected
  • The search API returns evidence chunks
  • Answers include citations

Validation: Treat uncited answers as failures and confirm the response includes source document titles.

AI Starter Prompt

I want to add mini RAG to an existing service. Please outline document intake, chunking, search API, cited responses, and failure policy in the smallest implementation order.
Read Concept Run Recipe
Day 3

3. Skill / Tool

(60-90 min) Repeated work

Turn repeated operating procedures into reusable skills with clear tool boundaries.

Outputs

SKILL.mdI/O contractvalidation command

Done Criteria

  • One repeated task is selected
  • SKILL.md includes trigger, inputs, procedure, and validation
  • Tool permission boundaries are defined

Validation: Start a new chat and confirm the procedure can be reproduced from the skill alone.

AI Starter Prompt

I want to turn one repeated project workflow into a Codex Skill. Please draft SKILL.md with trigger, inputs, procedure, files to read, validation commands, and recovery steps.
Read Concept Run Recipe
Day 4

4. Local LLM

(60-120 min) Model optionality

Use Ollama for quick local experiments and vLLM for server-style inference.

Outputs

base URL switchquality/cost tablerollback rule

Done Criteria

  • Hosted API vs local switching criteria are defined
  • Ollama or vLLM response is verified once
  • Quality/cost/ops table is written

Validation: Run the Day 1 harness with the same inputs and compare model behavior before and after the switch.

AI Starter Prompt

I want to compare Hosted API, Ollama, and vLLM for my AI feature. Please create a quality, cost, privacy, ops difficulty, rollback criteria table and a validation sequence.
Read Concept Run Recipe

Recommended Outputs

  • A minimum API such as /ai/ask
  • An eval baseline such as docs/ai-evals.md or evals/*.json
  • A RAG failure policy: answer unknown when evidence is missing
  • A model routing table: Hosted API, Ollama, or vLLM by scenario

Runnable Recipes

Make an Eval Harness Add Mini RAG Write a Project Skill Run a Local LLM

Last updated: June 22, 2026 · Version: v0.0.1

Send Feedback

Opens a new issue page with your message.

Open GitHub Issue