Dealing with useless AI answers

Generic AI output? Vague summaries? Hallucinations? The full troubleshooting flow for getting better AI answers.

Local AI is amazing. But sometimes the model gives you something generic, vague, or just plain weird. "This was a meaningful week with various activities." "You experienced many emotions." "It rained on Tuesday and that was important to you."

This page is the troubleshooting flow.

Step 1: Diagnose The Output

Before fixing anything, identify which failure mode you're seeing:

Symptom	Likely cause
Generic / platitudes	Model is too small, OR not enough input context
Hallucinations (made-up events)	Model is too small, OR prompt is too vague
Repetitive phrases	Model temperature too low (rare; defaults are fine)
Off-topic / random	Wrong model selected for the task
Empty / blank output	Ollama isn't running, or wrong model name
Lots of `<thinking>` tags showing	Response cleaning didn't catch your model's format — see Step 6
Cut off mid-sentence	Hit token limit — see Step 7

Step 2: Try A Better Model First

This fixes 80% of complaints.

The single biggest cause of vague AI output is using phi4-mini for synthesis tasks (weekly+ summaries, Q&A). It's a fast model, not a smart model.

Upgrade by going to Choosing Your AI Models and changing your weekly / monthly / Q&A models to:

phi4 (~9 GB) — strong upgrade
mistral:instruct (~4 GB) — different flavor, also good
llama3.1:8b (~5 GB) — solid all-rounder

Re-run the AI button. Most "useless AI answer" complaints disappear here.

Step 3: Check Your Input

Garbage in, garbage out. If your daily entry was:

"good day"

The AI literally has nothing to work with. The system prompt tells it to use specific events, but if there are none, it'll either invent some or shrug.

Fix: write a few sentences. The AI doesn't need novels — but two or three concrete things (what you did, who you saw, how you felt) are the minimum for good output.

Step 4: Tune The Prompt

If you've upgraded the model and your inputs are good but the output is still off, the prompt is the next lever.

Go to Customizing AI Prompts and try:

Adding "Be specific. Use only events explicitly mentioned." — fights hallucination
Setting word count limits — "Output exactly 80 words."
Asking for a specific structure — "Output 3 sentences: one about emotional shape, one about main events, one about what stood out."
Forbidding specific phrases — "Do not use the words 'meaningful', 'special', or 'important'."

Iterate one variable at a time.

Step 5: Adjust Chunking

If your weekly+ summaries are missing details from individual days, chunking is dropping context. Go to Chunking Settings Explained and:

Increase detailedWordCount (default 150 → try 250)
Turn on useDailyChunksForMonth if you haven't already
Verify maxTokensPerOperation matches your model's context window

Step 6: Output Has `<thinking>` Or Other Artifacts

Some models (especially DeepSeek-R1 variants) emit thinking tags before the answer. Time Garden's response cleaner removes the common ones, but if a new model has a new format, the artifacts can leak through.

Fix: switch to a non-reasoning model (e.g., phi4, mistral:instruct, llama3.1). Those don't produce thinking tags.

If you really want to use a reasoning model, you can manually edit the output, or wait for the cleaner to be updated in a future Time Garden version.

Step 7: Output Cuts Off

If summaries trail off mid-sentence, you've hit your model's output token limit. Two fixes:

Shorten what you ask for. A 5-sentence summary won't cut off; a 5-paragraph one might.
Use a model with a longer context window. Most modern models have plenty; if you're on something tiny, upgrade.

Step 8: Output Is Empty

If the AI button does nothing or returns blank output:

Verify Ollama is running. (Mac: llama icon in menu bar; Windows: Task Manager → Ollama.)
Verify the model is installed: ollama list in your terminal.
Check the model name in Time Garden settings exactly matches one from ollama list.
See AI not responding / Ollama issues for the deeper diagnostic flow.

Step 9: Q&A Specifically Feels Off

For Q&A, the answer quality depends on:

The question's specificity — "How was my month?" gets a generic recap; "What did I worry about that turned out fine?" gets gold
The chunking strategy for that period (see Chunking Settings Explained)
Whether Include Highlights / Ideas / Progress are toggled on (see Customizing AI Prompts)

Try rephrasing the question with a temporal anchor ("the first half of the month", "the week of the move") and watch quality jump.

Step 10: Last Resort — Just Edit It

The AI output is regular text in your note.

If the alias the AI generated is mostly right but one word feels off, just rewrite that word. The AI is a draft assistant, not a final judge. The journal is yours.

Many users let the AI run, then spend 30 seconds adjusting before moving on. That's a perfectly good workflow.

When To Just Skip AI Entirely

Some weeks the AI just won't get the vibe right. That's fine. Write the summary by hand. Set the rating from your gut. Skip Wheel of Life this week. Time Garden's structure works with or without Eternal — see What is Time Garden?.

Up Next

Choosing Your AI Models — the biggest quality lever
Customizing AI Prompts — finer tuning
AI not responding / Ollama issues — when the AI isn't responding at all

Dealing with useless AI answers

On this page