name: q-multimodal description: Extract visual, video, and audio features from media. Use for pixel features (Pillow), video frames (FFmpeg+Pillow), speech/audio features (openSMILE), music features (librosa), and visual semantic analysis (Gemini API batch or standard). metadata: version: "1.0.0"

Q-Multimodal

Multimodal media analysis: local low-level features (Pillow, openSMILE), mid/high-level visual semantic analysis (Gemini API). Local pipelines are fully generic and CLI-driven. Gemini pipelines are config-driven: copy scripts/gemini/pipeline_config.py to your project, customize, and run with --config <path>.

Setup (first time in a project)

Do this once when adopting the skill in a new project. The canonical layout is a target only for scripts/ and output/ — user assets (input data, media, .env, system prompt) can stay wherever they already live; point the scripts at them via absolute paths.

• Identify <BASE_DIR>: read the project's CLAUDE.md if it exists. If BASE_DIR isn't defined, ask the user which directory is the project root.
• Locate existing user assets — do not move them. Search, confirm each location with the user before proceeding:
• Input dataset file (xlsx/csv/json/parquet)
• Media directory (grouping structure — one subfolder per subject is ideal)
• .env with GOOGLE_API_KEY1-4 (Gemini only; check project root, home directory, common locations)
• System prompt file (Gemini only)
• Default: point at files in place. Set pipeline_config.py fields or CLI --input / --base-dir arguments to the absolute paths you found. Never move user data without explicit confirmation.
• Materialize only scripts/ and output/ under <BASE_DIR>. Copy the pipelines actually being used from ${SKILL_DIR}/scripts/ into <BASE_DIR>/scripts/:
• Local pipelines: pillow/, opensmile/, librosa/, common.py
• Gemini pipelines: gemini/batch/, gemini/standard/, gemini/pipeline_config.py (template → adapt in place or copy to <BASE_DIR>/scripts/pipeline_config.py)
• output/ is auto-created by scripts on first run
• Scan input columns (adapt reader to file format):

python -c "import pandas as pd; print(list(pd.read_<FORMAT>('INPUT', nrows=1).columns))"

• Confirm --id-cols with the user. The file column (--file-col) is always retained in output — every row identifies its exact media asset — and --id-cols adds further columns (e.g. a post id) carried through checkpoints and merges.
• Confirm --features with the user as well, so the feature categories a pipeline extracts match what the analysis needs.

References

Read the relevant reference file before executing a pipeline. These contain all flags, output column definitions, edge cases, and validation rules.

Local pipelines:

• references/image-visual-features.md — all feature categories, column definitions, computation notes
• references/video-visual-features.md — frame extraction, aggregation logic, dual output format
• references/audio-features.md — openSMILE feature sets, interpretable scores, feature levels
• references/music-features.md — librosa feature sets, tier-1 music scores, raw tonal/timbre block

Gemini pipelines:

• references/gemini-batch-workflow.md — full 6-step batch pipeline, retry workflow, error handling
• references/gemini-standard.md — standard pipeline details, model config, adapting for new projects
• references/multi-key-management.md — multi-key quota strategy, retry threshold decision table

Shared:

• references/checkpoint-format.md — column order, validation rules, output directory structure

Dependencies

Pipelines

Script path = ${SKILL_DIR}/scripts/<path>. Read the pipeline's reference file before running.

Local Pipelines (generic, CLI-driven)

librosa/music_features.py complements opensmile/audio_features.py: openSMILE covers speech/prosody, librosa covers music-native features (tempo, key/mode, harmony, timbre).

Shared utilities: common.py — read_input(), save_excel(), derive_subject(), merge_checkpoints()

Command pattern: python <script> --input <file> --base-dir <root> [--features ...] [--id-cols ...] [--subjects ...] [--preview] [--merge]

After a run finishes, check the output for rows with ok=False and confirm the failure count is zero or has an understood cause before treating the extraction as complete.

Gemini Pipelines (config-driven)

Both pipelines read a pipeline_config.py file that defines paths, schema, metadata formatting, and validation rules. Copy scripts/gemini/pipeline_config.py to your project and customize.

Standard (gemini/standard/gemini_standard.py): inline media, 25 workers, auto-retry. See gemini-standard.md.

Batch (gemini/batch/[0-5]*.py + utils.py): 6-step pipeline, 50% discount. URIs expire after 48 hours. See gemini-batch-workflow.md.

python 0uploadMedia.py --config /path/to/config.py --submit --max-batch-gb 2 --key 1
python 3checkStatus.py --config /path/to/config.py --poll
python 4retryErrors.py --config /path/to/config.py --preview
# >500 failures: batch retry
python 4retryErrors.py --config /path/to/config.py --submit
python 3checkStatus.py --config /path/to/config.py --poll
python 4retryErrors.py --config /path/to/config.py --collect
# <=500 failures or after batch retries: live fallback
python 4retryErrors.py --config /path/to/config.py --standard
python 5review.py --config /path/to/config.py --merge

Decision: >5 GB or >10 subjects and not time-sensitive → batch. Otherwise → standard. See gemini-standard.md.

Multi-key: Each GOOGLE_API_KEY{N} = 20 GB quota. See multi-key-management.md.

Adapting for New Projects

Local pipelines (Pillow, openSMILE, librosa): No modification needed. All project-specific values come from CLI args.

Gemini pipelines: Config-driven, no script modification needed. Scripts are copied to the project in step 2 above, then:

• Adapt <BASE_DIR>/scripts/pipeline_config.py (already copied from template in step 2)
• Set BASE_DIR, INPUT_PATH, SYSTEM_PROMPT_PATH to your project paths (SYSTEM_PROMPT_PATH relative to BASE_DIR, e.g., scripts/<prompt>.txt)
• Set GROUP_COL, FILE_COL, ANALYSIS_FIELDS to match your input schema and system prompt
• Implement subject_id() and format_metadata() for your domain
• Optionally implement validate_row() for field-specific validation rules
• Run any script with --config <BASE_DIR>/scripts/pipeline_config.py

Scope

Include: Image/video/audio feature extraction, Gemini visual semantic analysis, batch job management, checkpoint merging, multi-key quota management.

Exclude: ML model training, deep learning inference, real-time streaming analysis.