# Task recipes Use inspection for current fields, limits, price dimensions and model choices. There is no need to load every API guide for a supported recipe. These are the guided search/inspection workflows, not the full Primex offering; use the [Documentation map](docs.md) for other API capabilities. | Task | Operation filter | Selection decision | | --- | --- | --- | | Generate an image | `images.generate` | Prompt, dimensions, supported format and price units | | Read text aloud | `audio.speech` | Language, voice, length and supported controls | | Transcribe audio | `audio.transcribe` | Recording format and language coverage | | Extract scan/PDF text | `documents.ocr` | File format, pages and language coverage | | Extract typed fields or tables | `documents.analyze` | Invoice, receipt, identity or layout offering | | Translate text | `text.translate` | Target language and total input limit | ## Image or narration Search for the user's task and follow a match's inspection action. Fill its execution request with the actual prompt/text and requested supported controls. For speech, choose from the inspected voices; do not invent a voice ID. Use reported profile attributes; follow [voice selection](voices.md) and the optional inspection lookup when a required attribute is missing. A name alone does not establish gender. Respect an explicit voice preference and split long text only when that remains within the requested task and any budget. Deliver the saved image or MP3. ## Invoice or document Search for the document type. For typed invoice fields, choose invoice analysis; for plain text, choose OCR; for general tables, choose layout analysis. Language support for typed fields can differ from general OCR. Follow inspection's upload steps using the selected file's actual bytes, MIME and size. After successful finalization, use its READY `file_id` in the execution request. Return the requested fields/text and attach structured JSON if useful. ## Recording, then translation 1. Discover and inspect transcription. Upload the selected recording with the returned purpose and finalize it. Run transcription once. 2. Save the transcript. Discover and inspect translation, then use the transcript text and requested target language(s) as inputs to a new operation. 3. Deliver both outputs when requested. If translation fails, keep the completed transcript and report the remaining step; do not transcribe again. Each submission has its own operation UUID and receipt. Multiple target languages increase billable work. Use [Errors](errors.md) if either result is uncertain.