ai.KMITL Docs

Getting Better Results ​

Beyond writing great prompts, there are many strategies to get better, more accurate, and more useful results from ai.KMITL. This guide covers advanced techniques and best practices.

Understanding AI Behavior ​

How AI Models Work ​

AI models:

  • Predict text based on patterns in training data
  • Don't "know" things like humans do
  • Can make mistakes or "hallucinate"
  • Benefit from context and clear instructions
  • Improve with feedback and iteration

Key Insight

AI is a tool that amplifies your thinking - it works best when you guide it clearly and verify its outputs.

Iterative Refinement ​

The Iteration Process ​

Iteration is useful, but each new response uses quota:

  1. Start with a complete prompt: goal, relevant context, and desired format
  2. Review the response
  3. Edit the original text prompt if you left something out or asked unclearly
  4. Follow up when you genuinely need to build on the answer

Edit replaces the prompt and regenerates from that point, removing later messages. It does not refund the earlier response, and it is not available on messages with attachments.

Keep Iteration Focused

Every follow-up re-sends the conversation and its attached files. In a chat with a large document, plan related questions together and attach only relevant pages. For a new topic, start a new chat rather than carrying the whole thread forward.

Example Iteration ​

First attempt:

You: "Explain machine learning"

AI: [Generic explanation]

Edit the original prompt:

You: "Explain machine learning for a beginner, focusing on practical
applications rather than technical details. Include 3 everyday examples."

AI: [A focused explanation with examples]

Follow up if you need to build on it:

You: "How would the spam-filter example apply to my email app?"

AI: [Applies the explanation to your situation]

Providing Better Context ​

What to Include ​

For technical questions:

  • Your skill level (beginner/intermediate/advanced)
  • What you've already tried
  • Specific error messages
  • Your environment (Python 3.10, Windows, etc.)
  • What you want to accomplish

For creative tasks:

  • Target audience
  • Desired tone and style
  • Length requirements
  • Any specific elements to include/avoid
  • Examples of what you like

For learning:

  • Your current understanding
  • What confuses you specifically
  • How you learn best (examples, analogies, step-by-step)
  • What you need this knowledge for

Context Example ​

Weak context:

❌ "Help me with this code"

Strong context:

✅ "I'm a beginner learning Python. I'm trying to read a CSV file
and print each row. I've tried using the csv module but I'm getting
a 'FileNotFoundError'. I'm using Python 3.10 on Windows. Here's my code:
[code]

I want to understand both what's wrong and why."

Breaking Down Complex Problems ​

The Divide and Conquer Method ​

Instead of one huge request:

❌ "Build me a complete todo app with authentication, database,
and mobile-responsive design"

Break it into steps:

✅ Step 1: "Let's design the database schema for a todo app.
I need tables for users and todos. What fields should each have?"

✅ Step 2: "Great! Now help me write the SQL to create these tables"

✅ Step 3: "Now let's build the user authentication system..."

Benefits of Breaking Down ​

  • Easier to understand each piece
  • Better results per step
  • Can adjust direction as you go
  • Learn more from the process
  • Debug easier when issues arise

Using Web Search Effectively ​

Enable for:

  • ✅ Current events and news
  • ✅ Recent product releases
  • ✅ Latest technology versions
  • ✅ Real-time data (weather, stocks)
  • Fact-checking claims against current sources
  • ✅ Recent research papers

Don't need for:

  • ❌ General knowledge
  • ❌ Coding help (unless checking latest syntax)
  • ❌ Creative writing
  • ❌ Math problems
  • ❌ Historical facts

Making Web Search More Accurate ​

Be specific about time:

✅ "What AI announcements were made this week?"
✅ "Latest iPhone features (2024 model)"
✅ "Current Python version and release date"

Request sources:

✅ "Find recent statistics on climate change and cite your sources"
✅ "What do reputable news sources say about this topic?"

Verifying AI Responses ​

Always Verify For ​

  • Facts and statistics (especially without web search)
  • Code before running in production
  • Medical or legal advice (consult professionals!)
  • Financial decisions
  • Critical academic work

How to Verify ​

  1. Cross-reference with reliable sources
  2. Test code in a safe environment
  3. Ask for reasoning: "Why did you suggest this?"
  4. Use web search when verification needs current sources
  5. Consult experts for critical decisions

Critical Information

Never rely solely on AI for medical, legal, financial, or safety-critical decisions. Always consult qualified professionals.

Handling Hallucinations ​

What are Hallucinations? ​

When AI generates information that:

  • Sounds plausible but is incorrect
  • Includes made-up facts, citations, or statistics
  • Confidently presents wrong information

How to Reduce Hallucinations ​

1. Ask for sources:

✅ "Cite your sources for these statistics"
✅ "Where can I verify this information?"

2. Enable web search for current information:

Enable web search for recent facts that need up-to-date sources

3. Request uncertainty acknowledgment:

✅ "If you're not certain, please say so"
✅ "What's your confidence level in this answer?"

4. Verify critical information:

✅ Always double-check important facts
✅ Test code before using it

Red Flags ​

Watch out for:

  • 🚩 Overly specific statistics without sources
  • 🚩 Quotes from unverifiable sources
  • 🚩 Made-up technical terms
  • 🚩 Inconsistencies in the response
  • 🚩 Claims that sound too perfect

Choosing the Right Model ​

Choose by Usage Tier ​

Use the price marks in the model picker as a starting point:

TierGood starting tasks
$ — CheapSummarizing, translating, search, and everyday Q&A
$$ — MediumCoding and analysis that need more capability
$$$ — HighGenuinely hard problems a cheaper model has failed to solve

For search, enable the web search toggle only when the task needs current information. Use low reasoning effort for simple tasks when the selected model offers it.

When to Step Up ​

Start on a $ model. If the answer is not good enough, first check whether the prompt needs clearer instructions. To try a more capable model on the same question, choose it in the picker, then use Retry on the response. Keep using the cheaper model when it is working well.

Switching models in the same chat keeps the full conversation history, including files; it does not make a long thread cheaper by removing context. For large documents, use a short digest in a new chat instead.

See AI Models for model selection and Making Your Quota Last for the digest workflow and its limitations.

Using Files Effectively ​

Preparing Files ​

For images:

  • Use high resolution for text recognition
  • Crop to relevant areas
  • Ensure good lighting in photos
  • Rotate to correct orientation

For PDFs:

  • Keep content focused: a short excerpt uses less context than the whole document
  • Text PDFs work better than scans: scans are transcribed automatically, but a transcription can misread small print
  • Long PDFs are fine to attach: the AI receives an outline and reads the pages it needs, so mention the chapter or page when you know it
  • Remove unnecessary pages before uploading

For code:

  • Include complete context
  • Add comments explaining intent
  • Show error messages if debugging
  • Include relevant imports/dependencies

Files Count on Follow-Ups Too

Every message re-sends the conversation. PDFs are sent as converted text or an outline, but a file with Send full document turned on is sent whole again on each follow-up, even after switching models. Keep file-based chats focused; for premium analysis of a large document, the digest workflow in a new chat is still the cheapest option.

Good:

✅ "I've uploaded a diagram. Can you explain what it shows?"
✅ "This PDF contains my essay. Please check grammar and structure."
✅ "Here's a screenshot of an error. What might be wrong?"

Better:

✅ "I've uploaded a flowchart of my algorithm. Please:
1. Verify the logic is correct
2. Suggest optimizations
3. Identify potential edge cases"

Managing Conversations ​

When to Start a New Chat ​

Start fresh when:

  • Changing topics completely
  • Previous conversation was messy
  • Context is getting confusing
  • You want a "clean slate"
  • Hit model's context limit

Continue conversation when:

  • Building on previous topic
  • Need context from earlier
  • In middle of multi-step task
  • Model needs previous context

Organizing with Folders ​

Use folders to:

  • Group related conversations
  • Find chats easily later
  • Keep different topics separate
  • Archive completed projects

See Project Folders for details.

Quota Management ​

Making Messages Count ​

Strategies:

  1. Plan your question before sending
  2. Combine related questions into one message
  3. Use $ models for simple queries; speed alone does not tell you usage
  4. Use Edit to correct a text prompt instead of adding correction messages
  5. Keep context relevant, especially in chats with large attachments

When You're Running Low ​

  • Prioritize important questions
  • Use low reasoning effort for simple tasks
  • Enable web search only when you need current information
  • Use Continue, when offered, for a cut-off answer rather than Retry
  • Consider BYOK to use your own provider allowance instead of platform quota
  • Check both reset times: daily at 09:00 Bangkok time; weekly on Monday at 00:00 Bangkok time

See Making Your Quota Last for more.

Advanced Techniques ​

Chain Prompting ​

Build complex outputs through stages:

Step 1: "List the main features of a todo app"
→ AI lists features

Step 2: "For feature #3 (user authentication), design the database schema"
→ AI designs schema

Step 3: "Now write the Python code for this schema using SQLAlchemy"
→ AI writes code

Role Assignment ​

Have AI take specific roles:

✅ "Act as a senior software architect. Review my system design..."
✅ "You're a writing tutor. Help me improve this paragraph..."
✅ "Pretend you're a skeptical scientist. Critique this hypothesis..."

Constraint-Based Prompts ​

Add specific constraints:

✅ "Explain quantum computing in exactly 3 sentences"
✅ "Write code using only standard library (no external packages)"
✅ "Describe this concept using only everyday analogies"

Few-Shot Learning ​

Provide examples of what you want:

✅ "Here are 3 examples of the format I want:
Example 1: [your example]
Example 2: [your example]
Example 3: [your example]

Now create one for: [new item]"

Common Pitfalls ​

1. Not Providing Enough Information ​

Problem:

❌ "This doesn't work"

Solution:

✅ "This Python function returns None instead of the sum.
Here's the code: [code]. Expected: 15, Got: None"

2. Expecting Perfection First Try ​

Problem: Getting frustrated when first response isn't perfect

Solution: Refine and iterate. Use Edit to fix a text prompt; use follow-ups to build on an answer. Each regeneration or follow-up uses quota.

3. Not Verifying Critical Information ​

Problem: Using generated code or facts without checking

Solution: Always test, verify, and validate important outputs

4. Ignoring Model Limitations ​

Problem: Expecting AI to know everything or be always right

Solution: Understand limitations, verify facts, and use web search when you need current sources.

5. Not Learning from Interactions ​

Problem: Repeating same unclear prompts

Solution: Note what works, refine your approach, build prompt templates

Quality Checklist ​

Before sending your message, ask:

After receiving a response:

Troubleshooting Common Issues ​

Response is Too Generic ​

Fix:

  • Add more specific details
  • Provide examples of what you want
  • Specify the format
  • Add constraints

Response is Wrong ​

Fix:

  • Enable web search if the answer needs current information
  • Edit an unclear text prompt to include the correct information
  • If a clear prompt still gets a poor answer, choose a more capable model and use Retry
  • Verify your own assumptions

Response is Too Technical ​

Fix:

  • Specify your skill level
  • Ask for simpler language
  • Request analogies
  • Ask "explain like I'm 5"

Response is Too Simple ​

Fix:

  • Ask for more detail
  • Request technical depth
  • Ask about edge cases
  • Request examples

AI Seems Stuck ​

Fix:

  • Rephrase your question
  • Start a new conversation
  • Try a different model
  • Break problem into smaller pieces

Best Practices Summary ​

Do's ✅ ​

  1. Iterate on responses
  2. Provide relevant context, not entire files when a few pages will do
  3. Break down complex tasks
  4. Verify critical information
  5. Choose appropriate models
  6. Organize with folders
  7. Learn from each interaction
  8. Experiment with techniques

Don'ts ❌ ​

  1. Don't expect perfection immediately
  2. Don't skip verification
  3. Don't use AI for critical decisions alone
  4. Don't forget to provide context
  5. Don't waste quota on unclear prompts
  6. Don't ignore model strengths
  7. Don't give up after one try
  8. Don't forget AI has limitations

Continuous Improvement ​

Learn from Each Session ​

  • Note what prompts worked well
  • Save useful conversation patterns
  • Develop your own prompt templates
  • Understand each model's style
  • Build a personal playbook

Experiment Regularly ​

  • Try different models for same task
  • Test various prompt styles
  • Compare results
  • Refine your approach
  • Share learnings with others

The Path to Mastery

Getting great results is a skill that improves with practice. Every conversation teaches you something new about effective prompting and AI interaction.

Next Steps ​

Made with ❤️ by KDMC (KMITL Data Management Center)

Copyright © 2025 KMITL