Agent Tool Design Guide
Design tool interfaces agents can route, call, and recover from reliably.
The skill recommends sandboxing and acknowledges safety constraints, with no visible malware, credential theft, or covert exfiltration. However, least privilege, user confirmation, sensitive-data flow, external side effects, and rollback are not specified end to end, so points are deducted.
The instructions are internally coherent and include recovery guidance, a testing checklist, and a readable Python utility. The utility has no committed tests, uses crude heuristic checks, and the key design claims are not statically reproduced; the score is therefore capped and reduced for thin edge-case and failure handling.
Activation triggers, scope, and boundaries with adjacent skills are clearly stated, and the skill is presented as platform-agnostic. It lacks Chinese-language guidance, fully worked input/output contract coverage, and broader non-fit boundaries, so points are deducted.
The skill has a clear information architecture, progressive-disclosure references, examples, version and update metadata, and repository-level MIT licensing. It lacks an explicit maintainer, changelog, installation/dependency notes, and a clear update path; cited case material is also not fully attributed in the supplied evidence.
The material provides directly usable design structures, naming guidance, recovery patterns, checklists, and a generator utility. The generator is not a complete runtime or schema validator, examples and case coverage are limited, and practical recommendations still require system-specific validation; the static ceiling therefore applies.
The supplied material includes auditable documentation, source code, and quantitative case-study claims. It does not include committed tests or independent reproduction for this skill, and some evidence is represented only by names or placeholder claim identifiers, so corroboration is weak.
- Use the “arbitrary bash command” pattern only with strict isolation, least privilege, resource limits, auditing, and human approval; do not grant direct access to production hosts or sensitive data.
- The case-study figures for success rate, latency, and token use lack reproducible experimental details in the supplied material and should not be treated as general guarantees.
- MCP naming examples still need validation against the target deployment, version, and server constraints; generated descriptions and error templates should be tested against real schemas, abnormal inputs, and API behavior.
What does this skill do, and when should you use it?
This Agent Skill focuses on the tool-interface layer of an agent system, covering descriptions, schemas, response formats, naming, error recovery, and MCP tool design. It treats each tool as an explicit contract between deterministic software and a non-deterministic agent, using consolidation and namespacing to reduce routing ambiguity. It also addresses response-size choices, primitive tools, tool testing, and iterative improvement from observed failures. It fits engineers building or auditing agent tool catalogs, but does not own project-level pipeline architecture or decisions about splitting work among sub-agents.
It provides concrete methods for designing and auditing individual tools or tool catalogs: descriptions should state what a tool does, when to use it, what inputs it accepts, and what it returns; parameters should include types, constraints, defaults, and examples; names should follow verb-noun conventions with consistent fields and fully qualified MCP names in the form ServerName:tool_name. It guides actionable error messages, concise and detailed response modes, and decisions about merging overlapping tools. It also includes an audit checklist, five testing criteria, and a Python-style pseudocode pattern for analyzing tool failures and improving descriptions.
- An agent repeatedly selects the wrong tool, so an engineer uses the skill to identify overlap and restructure the catalog.
- A team is designing a new API or MCP tool and needs a clear schema, return shape, and recovery path.
- Tool calls frequently use malformed inputs, so a developer adds format examples, defaults, and explicit constraints.
- A tool catalog is growing rapidly, so an architect evaluates namespacing, consolidation, and primitive capabilities.
- Tool responses consume too much context, so an engineer designs concise and detailed response modes.
What are this skill's strengths and limitations?
- Covers the full interface layer: descriptions, schemas, responses, naming, recovery, and catalog governance.
- Includes a practical audit checklist and five evaluation criteria: unambiguity, completeness, recoverability, efficiency, and consistency.
- Clearly separates tool-interface work from project architecture and multi-agent decomposition.
- Offers specific guidance on consolidation, response formats, and fully qualified MCP naming.
- It is design guidance, not a runnable tool implementation or framework adapter.
- The code examples are primarily Python pseudocode, and no dedicated automated test suite for this skill is provided.
- MCP, filesystem, and Unix-tool material describes design patterns; deployment still requires your own runtime, permissions, and implementation.
- Some production results are workload-specific case evidence, and the skill recommends reproducing them on the target catalog.
How do you install this skill?
After obtaining the repository, copy the complete tool-design directory into the target Agent Skills directory while preserving SKILL.md and the folder structure. For example: mkdir -p .codex/skills && cp -R skills/tool-design .codex/skills/. It can also be copied to .claude/skills/ or .agents/skills/. Do not flatten SKILL.md into a single file. The README also supports using the repository as a plugin directory, but it does not document client-specific standalone installation steps beyond these layouts.
How do you use this skill?
In an Agent Skills-compatible client, describe the work as a tool-interface task, for example: “Design a clear MCP tool description, parameter schema, response format, and actionable error messages, then check the catalog for overlapping tools.” Activate it when the unit of work is one tool or a tool set. Route project pipeline architecture, sub-agent decisions, and trajectory-level context optimization elsewhere.