---

name: files-and-multimodal description: "Guides agents handling CrewAI crewai-files, input_files, file source resolution, multimodal agents, provider file constraints, upload caches, and file or document tool adjacency." disable-model-invocation: true metadata: disco-role: operating license: MIT

Files and Multimodal

Use this sub-skill when a task involves CrewAI file inputs, crewai-files, images, PDFs, audio, video, text files, input_files, multimodal agents, provider file formatting, upload behavior, or deciding whether to use native file inputs versus file/document tools.

Route First

• For File, ImageFile, PDFFile, AudioFile, VideoFile, TextFile, FilePath, FileUrl, FileBytes, FileResolver, input_files, and prompt key patterns, read references/file-inputs.md.
• For provider support, MIME/type limits, OpenAI Responses versus completions behavior, Bedrock/Gemini/Anthropic formatting, and file handling modes, read references/provider-constraints.md.
• For inline versus URL versus upload delivery, prefer_upload, UploadCache, cleanup helpers, and safe cache expectations, read references/uploaders-and-cache.md.
• For missing optional dependencies, unsupported MIME types, URL/path confusion, base64 misuse, upload cleanup, and multimodal flag migration, read references/troubleshooting.md.
• To inspect local paths, safe URL metadata, data URIs, base64 inputs, and provider constraints without LLM calls, uploads, credentials, or network fetches, run scripts/check_file_inputs.py with --help first.

Boundaries

• Stay here for native CrewAI file inputs, crewai-files classes, file source coercion, provider file constraints, multimodal Agent/Crew/Task input wiring, upload cache behavior, and adjacent file/document tool decisions.
• Use ../tools-and-mcp/SKILL.md for official crewai_tools document tools such as FileReadTool, DirectoryReadTool, PDFSearchTool, CSVSearchTool, JSONSearchTool, OCR, custom tools, and MCP tool routing.
• Use ../memory-knowledge-and-rag/SKILL.md for knowledge sources, RAG loaders, embeddings, vector stores, and retrieval/search architectures built on documents.
• Use ../llm-and-providers/SKILL.md for provider credentials, model selection, base URLs, API variants, streaming, and general LLM compatibility beyond file payload constraints.
• Use ../core-runtime/SKILL.md for ordinary Agent, Task, Crew, process, kickoff, callback, guardrail, and output design when files are not the central issue.
• Return to ../../SKILL.md when a request spans CrewAI project routing, CLI setup, tools, RAG, providers, and file inputs together.

Safe Defaults

• Prefer native input_files with typed crewai_files objects for model-readable file payloads; prefer file/document tools when the agent should browse or search local documents during tool use.
• Validate MIME type, provider support, and size limits before a real kickoff; small local fixtures are safer than large PDFs, videos, or remote URLs during debugging.
• Treat uploads as side-effectful provider operations; do not enable prefer_upload=True unless the user has approved credentials, retention assumptions, and cleanup needs.
• For legacy image URL/path tasks that rely on multimodal=True, prefer migrating to native input_files where possible; keep multimodal=True only when the intended behavior is the built-in image tool path.

Usability Targets

• Validate mixed local path, URL, and data-URI/base64 inputs for OpenAI Responses versus Bedrock constraints without triggering network, upload, or LLM calls.
• Diagnose why MIME detection falls back to application/octet-stream, why a provider rejects a PDF/audio/video file, or why a non-native file path is handled by a file tool instead of the LLM payload.