Testing Prompts and Responses of Free LLMs

[chinaexpert1]

Overview

Find and test prompts against free LLMs, so that we can know if that is a viable option for our pilot use

Action Items

• find a list of Free LLMs to use, and add details of how to access in the comments (e.g., XYZ LLM is available via API, ABC LLM requires hosting it ourselves, etc.)
• create a document in the drive with findings

Resources/Instructions

• ... Here is the second prompt experiments for website issue #1407 aand #1445 covering all Claude current models

Successful Prompt Used on Claude Sonnet 4.5

You will have to edit this to include the issue comments

Analyze the GitHub issue comments below and create STAR-style resume bullets for ALL contributors who completed tasks on this issue.
[ISSUE COMMENTS HERE]
For each contributor:
Identify their username from the comment history
List each specific task they completed (do not group tasks together)
Write one STAR-style resume bullet for each task completed
Format your response as:
[Username] • [STAR bullet for task 1] • [STAR bullet for task 2] • [STAR bullet for task 3]
[Username] • [STAR bullet for task 1] • [STAR bullet for task 2]
Requirements:
Do NOT assume there is only one contributor
Do NOT attribute work to "group effort" or "the team"
Do NOT include introductory text or closing questions
Do NOT summarize multiple tasks into one bullet
Count all discrete tasks mentioned in the issue and create separate bullets for each
Ignore any images in the input

• you can find a txt file for both issue comments to test and the winning Claude prompt here

[chinaexpert1]

Please provide update @sandy3w

Also Sandy please ask each LLM you ask for the best prompt for them BASED on the Claude prompt, they know themselves best.

Instructions

1. Progress: "What is the current status of your project? What have you completed and what is left to do?"
2. Blockers: "Difficulties or errors encountered."
3. Availability: "How much time will you have this week to work on this issue?"
4. ETA: "When do you expect this issue to be completed?"
5. Pictures (if necessary): "Add any pictures that will help illustrate what you are working on."

You can use this template

1. Progress: 
2. Blockers: 
3. Availability:
4. ETA:
5. Pictures (if necessary): 

[chinaexpert1]

I assigned myself I've done half of it.

[chinaexpert1]

Assigned Sarish. He has taken on the free API task these tasks probably should be merged.

[chinaexpert1]

Good to know...free LLMs we can use instead of paying high ChatGPT token budgets:

OpenRouter — GUI: https://openrouter.ai/chat — Many models, solid limits
Google AI Studio (Gemini) — GUI: https://ai.google.dev/gemini — Excellent structured extraction
Mistral (La Plateforme) — GUI: https://console.mistral.ai/playground — Strong for labeling & embedding
Cerebras — GUI: https://inference.cerebras.ai/playground — Very high token limits
Groq Console — GUI: https://console.groq.com/playground — Fast & accurate extraction
Cohere — GUI: https://dashboard.cohere.com/playground — Good for embeddings & classification
HuggingFace Model Chat — Chat UIs per model — Large OSS model selection
Cloudflare Workers AI — GUI: https://developers.cloudflare.com/workers-ai/playground — Good free models
AI21 Studio — GUI: https://studio.ai21.com/ — Jamba excels at structured output
Upstage Solar — GUI: https://console.upstage.ai/ — Very strong for text extraction
Hyperbolic Playground — GUI: https://app.hyperbolic.xyz/chat — Great for Qwen/Llama
SambaNova Cloud — GUI: https://cloud.sambanova.ai/ — Large high-quality models

[christopher0331]

Sharing a few prompts that are easy to run against free LLMs (ChatGPT free tier, Claude free, Gemini) for your testing matrix — they give clearly gradable outputs, which makes model comparison fair:

1. Structured rewriter (good for measuring instruction-following):
   "Rewrite the text below to a 6th-grade reading level. Keep all facts. Output ONLY the rewritten text. Text: [paste]"

2. First-principles explainer (measures reasoning + analogy quality):
   "Explain [topic] to someone smart but new to the field. One real-world analogy, then three layered examples."

3. Brutal editor (measures concision + measurable outcomes):
   "Rewrite each bullet to start with a strong verb and end with a measurable result. Flag anything vague. Bullets: [paste]"

4. Debugging partner (measures stepwise reasoning):
   "List the 3 most likely root causes in plain English, then the minimal fix. Error: [paste]"

Tip: score each on a 1-5 rubric (accuracy, format adherence, concision) so free vs paid models are comparable. I keep a free starter set of these in a public repo (christopher0331/ai-prompt-samples) if you want the full list to seed your tests.