Skip to content
Back to skills

Files And Multimodal

ASecurity

"Guides agents handling CrewAI crewai-files, input_files, file

  • 247 stars
  • 0 votes
  • 0 copies
  • 2 views
  • Added September 8, 2026
ai-agentsdebuggingapi

Works with

  • cli
  • api
  • mcp

Security analysis

A100/100

Pro scans all 6 files and shows the line behind each finding

Scanned September 8, 2026

npx -y skills add VectorSpaceLab/AREX-Skill --skill files-and-multimodal --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Files And Multimodal?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Files And Multimodal
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/vectorspacelab-files-and-multimodal/badge)](https://www.skillsdirectory.com/skills/vectorspacelab-files-and-multimodal)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
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](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](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](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](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](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](../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](../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](../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](../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](../../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.

Files in this skill

  • SKILL.md4 KB
  • references/file-inputs.md7.4 KB
  • references/provider-constraints.md7.1 KB
  • references/troubleshooting.md7.9 KB
  • references/uploaders-and-cache.md6.6 KB
  • scripts/check_file_inputs.py20.1 KB

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.