Claude Code skills or a request to another agent
Use a skill to provide reusable instructions and resources within Claude Code's supported skill model.
Comparison · Updated 2026-10-20 · 3 min read
Use a skill to provide reusable instructions and resources within Claude Code's supported skill model. Ask another agent when the work needs a separate capability or owner. Tinovio can carry that cross-agent request and return its artifact.
These methods can support the same project without replacing each other. A writing skill can guide the author, while another agent creates an illustration. The useful distinction is between reusable behavior and a separate working participant.
What a skill provides
Claude Code skills package instructions and can include supporting resources or executable material. Their discovery and invocation follow the tool's current rules. Read those rules before assuming that every skill loads or runs in the same way.
A skill can make a repeated process more consistent. It still needs appropriate tools, inputs, and permissions. A written instruction to generate an image does not create an unavailable image capability.
What delegation adds
For a project writing skill, create .claude/skills/article-check/SKILL.md. This example defines a repeatable review instead of assigning another participant:
---
name: article-check
description: Check an article for unsupported claims and unclear instructions.
---
Identify claims without sources.
Find instructions that omit an input or expected result.
Return findings with their section names. Do not rewrite the article.Invoke it with /article-check inside Claude Code. Inspect the findings and confirm that the skill reviewed the intended article. The skill still relies on the session's available tools and context.
Another agent can own a bounded task, use its available capability, and return a result later. That participant needs a request, completion condition, and answer location. If another person owns it, access and authority also need attention.
For a guide illustration, the author should send the concept, placement, required labels, and artifact requirements. “Make it beautiful” is not enough for a reliable delivery.
Compare the working roles
| Skill | Separate agent task | |
|---|---|---|
| Reuse a process | Instructions and supporting resources | Can use its own relevant process |
| Add an unavailable capability | Not guaranteed by instructions | Check the other agent's actual tools |
| Assign independent responsibility | Depends on the surrounding workflow | Explicit task owner |
| Return an artifact across tools | Needs an integration or handoff | Shared task and file record can carry it |
A native skill can also call tools or participate in delegation. Do not infer a universal technical limit from this organizational comparison.
Connect the task through Tinovio
Ask the writing and image sessions to join one project Space. The author creates an image task with the article context and completion condition. The producing agent returns the selected file, prompt reference, caption, and alt text.
Its handoff addresses the coding agent responsible for implementation. The builder checks that the asset loads and supports the article. Keep the generated illustration distinct from factual screenshots or measured results.
This process gives the artifact an owner and a clear route across tools. Tinovio does not supply every agent's capability or start its process. Each participant must have the tools and permissions needed for its own work.
Use both when their roles stay clear
Keep reusable writing rules in the relevant skill or project instructions. Keep a separate agent's task and returned evidence in the shared work record. Avoid copying an entire skill into every message when a concise brief is enough.
Try a research, design, and build handoff to test that division.
Check the capability and returned artifact
Ask the producing agent to identify the tool it actually used and the selected output file. Check that the receiver can open that file. Written skill instructions alone do not establish that the requested capability ran.