Ollama prompt experiment

ollama-try sends one prompt to a local Ollama model and prints the answer. Use it to see whether a model will return scene files without a plan or a chain of thought. It does not write a film folder, and it does not call ComfyUI.

The program is quickies/comfyui/ollama-try. From ~/Projects/short-films:

./quickies/comfyui/ollama-try --help
./quickies/comfyui/ollama-try --self-test

--self-test only checks that <think> blocks are removed. It does not call Ollama.

What you need

Ollama is running at http://127.0.0.1:11434.
The GPU is free. This machine has one GPU. Do not run this while ComfyUI, MiniMax H3, or IndexTTS is using it.

Check both:

curl -sf http://127.0.0.1:11434/api/version
curl -sf http://127.0.0.1:11434/api/ps
nvidia-smi --query-gpu=memory.used,utilization.gpu --format=csv
curl -sf http://127.0.0.1:9000/queue

/api/ps should show no model, or only the model you are about to test. ComfyUI's queue should have nothing running and nothing pending. Video memory should be a few GB, not the 16 GB a loaded MiniMax job holds.

If the queue is empty and ComfyUI is still holding the weights, release that cache, then look at the memory line again:

curl -sf -X POST http://127.0.0.1:9000/free \
  -H 'Content-Type: application/json' \
  -d '{"unload_models": true, "free_memory": true}'

The next ComfyUI job will load the weights again.

Where the files go

Put every run in its own folder under quickies/comfyui/experiments/. The answer is written there with --out. Thinking, when you ask to see it, is stderr, so redirect stderr into the same folder. Do not send the answer to /tmp. That path is outside the project and does not show up in the workspace.

quickies/comfyui/experiments/1978-lata-vikram/
  system.txt    instructions: format, locks, no commentary
  story.txt     the high-level story only
  answer.txt    stdout, the model's files
  stderr.txt    empty when no thinking was returned

1978-lata-vikram/ is the run already on disk. answer.txt is the model's text. system.txt and story.txt are what was sent.

The call that matches that run

qwen-3.8-27B-Q4XL is the model for a full set of clip files. Its Modelfile template is only {{ .Prompt }}, so a system role may never reach it. --preface-system copies system.txt into the front of the user turn. --generate calls /api/generate, which matches that template. Thinking stays off unless you pass --think.

cd ~/Projects/short-films

./quickies/comfyui/ollama-try \
  --model qwen-3.8-27B-Q4XL \
  --temperature 0 \
  --num-predict 4096 \
  --show-thinking \
  --generate \
  --preface-system \
  --system quickies/comfyui/experiments/1978-lata-vikram/system.txt \
  --prompt-file quickies/comfyui/experiments/1978-lata-vikram/story.txt \
  --out quickies/comfyui/experiments/1978-lata-vikram/answer.txt \
  2> quickies/comfyui/experiments/1978-lata-vikram/stderr.txt

--num-predict 4096 is the token cap for a four-clip answer. A longer film needs a higher cap. --temperature 0 keeps the run repeatable. --show-thinking prints any reasoning on stderr. With thinking off, stderr.txt should be empty.

Stdout is also printed in the terminal. --out writes the same text again.

When the run finishes, unload the model so the GPU is free for ComfyUI:

curl -sf http://127.0.0.1:11434/api/generate \
  -H 'Content-Type: application/json' \
  -d '{"model":"qwen-3.8-27B-Q4XL","keep_alive":0}'

A new story

Copy the folder, replace story.txt, keep system.txt unless you are testing a different instruction, and point --out and the stderr redirect at the new folder.

mkdir -p quickies/comfyui/experiments/my-run
cp quickies/comfyui/experiments/1978-lata-vikram/system.txt \
   quickies/comfyui/experiments/my-run/system.txt

Write quickies/comfyui/experiments/my-run/story.txt, then run the same command with my-run in place of 1978-lata-vikram.

What to look at afterward

Open answer.txt. For this experiment a usable answer has only the marked sections ### arc.md and ### clip-01-15s.txt onward. No plan before them.

Then check the pictures against the headers:

Each lock in arc.md appears again in the clip, character for character.
Shot 1 begins with the style sentence, in brackets: [Shot 1].
overall_soundscape and non_diegetic_music: N/A are present.
handoff-from: clip-NN last frame begins on the picture named in the previous end-frame.
A new room, a time jump, or a costume change uses handoff-from: none.
Spoken lines are only the ones the story asked for.

The 1978 answer is the sample of a clean channel and a weak film. stderr.txt is empty, and answer.txt has no commentary. The locks are paraphrased, the handoff pictures do not continue, and the sound lines are missing. Those misses are recorded in quickies/comfyui/story-pipeline.md.

Other switches

Switch	Effect
`--think`	Lets the model reason. Leave it off for this experiment.
`--show-thinking`	Prints a `thinking` field, or a removed `<think>` block, on stderr.
`--raw`	Leaves `<think>` blocks in the answer.
`--model mag-mel-R1-12B`	Tries a smaller model. Do not treat it as able to hold a full set of locks.
(no `--generate`)	Calls `/api/chat` instead. Use this when you want to see whether a system role survives.
`--host`	Another Ollama address. The default is `http://127.0.0.1:11434`.