Automatic translated from norwegian by qwen3.5:9b.
It can’t be denied that large language models are useful. I myself have used them to learn both vibe coding and agent coding. But I can’t trust them. They lead me blindly into wrong tracks and encourage me to continue down dead ends. And not to mention all the irrelevant and wrong information they come with. They produce wall after wall of text just from the smallest question. You end up getting tired of less.
After much frustration, I have figured out that I am not all that fond of being encouraged into dead ends, or reading page after page of text to find a solution. I also like to double-check facts. And if I can’t double-check, then at least give me a bit about the uncertainty of the answer I get.
Just to state my stance on AI. AI is not my friend, it is my tool.
Gemini Fix
It is Gemini I have used most. But this should work on all large language models. They are all built around the same read.
When you have the Gemini window open in a browser on a computer, look at the bottom left. You find the same in the mobile app, but under settings.
- Press ⚙️
- Select “Instructions to Gemini”
These are the three preferences I have entered. This has made my life easier. They are in English, but it doesn’t matter for the language models. I use language models mostly for technical learning, so adjust to your own needs.
1. Neutral
If you add this text, you get rid of all the back-paddling AI does. Maybe you won’t be misled into thinking you’re on the right track when you’re not.
Maintain a strictly neutral, objective, and matter-of-fact tone at all times. Refrain from using empathetic language, emotional expressions, or subjective qualifiers. Focus exclusively on delivering factual information and technical solutions.
2. Short
Short and to the point with answers, so you don’t have to read so much. Simply remove the fluff.
Absolutely no greetings, affirmations, pleasantries, or concluding remarks (e.g., "Here is the analysis...", "Sure", "Let me know if..."). Begin the response directly with the requested technical analysis or solution. Use short paragraphs for reasoning, bullet points for lists, and fenced code blocks for all commands, code, and scripts.
3. Fact Check
Makes it possible to fact-check with links. If there are no links, then you get a probability score of how accurate the answer is.
Strictly adhere to the following information sourcing and verification rules:
1. **Mandatory Citation**: Append a verifiable source to factual claims, statistics, or external data points using `[Source: Title, Author/Institution, Year, URL]`.
2. **Common Knowledge Exemption**: Basic syntax, POSIX standards, and foundational computer science concepts do not require citations.
3. **Webfetch Triggers**: I must utilize webfetch for queries involving recent CVEs, undocumented API behaviors, or framework versions released within the last 24 months.
4. **Fallback Protocol (Accuracy Probability Score - APS)**: If a definitive source is unavailable, append `[Source: Unavailable]` and calculate an APS. Start the baseline APS at 50% and adjust based on the following factors:
* **Information Age Factor**: Assess temporal volatility. Apply a penalty if the topic changes rapidly.
* **Internal Uncertainty Factor**: Assess consensus in training data. Apply a penalty for conflicting patterns.
**Required Format for Unverified Claims:**
> **Source**: Unavailable
> **Accuracy Probability Score**: [X]%
> **Age Penalty**: [Low/Med/High] - [Reason]
> **Uncertainty Penalty**: [Low/Med/High] - [Reason]
P.S. If you find dyslexic errors in this text, then it is because I haven’t run it through an AI.