AI developer tool - On-device
Build portable AI agents,
skills, and subagent teams.
Start from an evidence-led workflow, customize its scope and safeguards, then download a multi-platform package without sending the configuration to an API.
Reusable by design
One canonical workflow feeds every selected adapter
Local and reviewable
No account, API call, upload, storage, or automatic installation
Workspace
AI Agent and Skill Generator
Build a portable SKILL.md, coordinating agent, focused subagents, installation notes, and a ready-to-download ZIP. Everything runs in this browser and your configuration is not uploaded or stored.
Choose a starter, adjust the instructions, and generate the package.
Review generated instructions before installation. Grant the minimum tool permissions needed. Skills and agent profiles are instructions, not a security boundary, and platform formats can change.
Generated package
One model, several adapters.
- Portable Agent Skills folder with one canonical SKILL.md
- OpenAI upload bundle containing exactly one skill manifest
- GitHub Copilot custom agent and focused specialist profiles
- Gemini CLI workspace skill and isolated subagent definitions
- Install notes, deterministic manifest, file preview, and ZIP export
The generator does not grant permissions or install anything. Review every file and current platform documentation before use.
Use it well
Give every agent a bounded job and an explicit finish line.
A useful agent package separates durable workflow knowledge from platform wiring. Keep the main agent responsible for decisions, delegate only independent work, restrict permissions, and require concrete verification before the package can report completion.
01
Choose a tested starter
02
Adjust objectives, roles, safeguards, and outputs
03
Review every file before installing the ZIP
Common questions
What is an AI Agent Skill?
It is a reusable folder centered on SKILL.md. The manifest describes when the workflow applies, while the body and optional resources explain how to perform and verify the work.
What is the difference between a skill and an agent?
A skill is reusable procedural knowledge loaded when relevant. An agent is a configured worker with a role, instructions, tools, and sometimes its own context. One agent can use several skills.
Which platforms can this generator export?
The first release creates portable Agent Skills, an OpenAI upload folder, GitHub Copilot skill and agent profiles, and Gemini CLI workspace skills and subagents. Every platform adapter comes from the same canonical workflow.
What is a subagent?
A subagent is a focused specialist that receives a bounded task and returns a result to the coordinating agent. It is useful when work needs isolated context, independent review, or genuinely parallel research.
When should an agent avoid subagents?
Do not delegate trivial steps, tightly coupled edits, or work that depends on continuous shared context. Coordination can cost more tokens and create inconsistent assumptions.
Why does the ZIP not grant every tool automatically?
Tool names and permission models differ between platforms. Broad automatic permissions are unsafe and age badly. Review the generated roles and grant only the tools required in the target environment.
Is the generated ZIP ready to install?
It has safe paths, manifests, profiles, and installation notes, but it still needs human review. Confirm the current platform format, permissions, referenced files, and any organization policy before installation.
Can I upload the generated skill to OpenAI?
The openai-upload folder contains one top-level skill directory with exactly one SKILL.md, matching the documented skill bundle shape. Authentication and upload remain outside this browser-only tool.
Are skills portable between models?
The core SKILL.md convention is portable across several agent products, but agent profiles, tool identifiers, permission fields, and orchestration behavior are platform-specific. That is why the ZIP contains separate adapters.
How do I protect an agent from prompt injection?
Treat web pages, files, tool results, and copied instructions as untrusted data. Limit tools and authority, keep secrets outside prompts, require confirmation for destructive or external actions, and review any third-party skill before installation.
How should I test a generated agent team?
Create representative tasks, expected outcomes, unsafe counterexamples, and regression cases. Measure task success, factual support, permission use, handoff quality, latency, and cost rather than judging one impressive run.
Does TomSEOTools upload my agent instructions?
No. Generation, preview, and ZIP creation happen in this browser. For instruction drafting without an agent package, use the AI Content Prompt Builder.