How to use the AI-First Engineering Practices agent skill

What does the ai-first-engineering skill make the agent do?

The skill is advisory text. It shapes how the agent thinks and writes when you ask it to design a process, review code or set test standards. It does not run anything on its own, and it does not call tools or the network.

The source file sets out three shifts in process: planning quality matters more than typing speed, evaluation coverage matters more than personal confidence, and review focus moves from syntax to system behaviour.

• Architecture: prefer explicit boundaries, stable contracts, typed interfaces and deterministic tests, and avoid implicit behaviour spread across hidden conventions.
• Code review: focus on behavioural regressions, security assumptions, data integrity, failure handling and rollout safety, and spend little time on style issues that automation already covers.
• Hiring and evaluation: look for people who break down ambiguous work, define measurable acceptance criteria, write high-signal prompts and evaluations, and hold risk controls under deadline pressure.
• Testing: require regression coverage for the domain being changed, explicit edge-case assertions and integration checks at interface boundaries.

How do you install the skill?

The skill is a single file, SKILL.md. You can install it in one of two ways.

Option one is to download the ZIP from the skill's page on this directory and unzip it into your personal skills folder, ~/.claude/skills/, or into a project's .claude/skills/ folder if you want it only for one repository.

Option two is to upload the ZIP in the Claude apps under Settings → Capabilities → Skills.

• Personal install: unzip into ~/.claude/skills/ so the skill is available across your projects.
• Project install: unzip into the .claude/skills/ folder inside a specific repository.
• App install: upload the ZIP under Settings → Capabilities → Skills.

How do you trigger the skill with a natural request?

You do not need a special command. Describe the task in plain language that matches the skill's purpose: process design, architecture, code review or test standards for a team that uses AI to write code.

The skill's description is written in Japanese, and its guidance is also in Japanese. Writing your request in Japanese will match the source most closely, though the agent should still be able to apply the principles to an English request.

• Review a pull request for behavioural regressions and rollout risk rather than style.
• Draft test standards for code generated by an AI agent in our service.
• Suggest an architecture with explicit boundaries and typed interfaces for a team that ships AI-written code.
• Set measurable acceptance criteria for a backlog of ambiguous tickets.

What examples show it working?

The skill does not include templates, commands or sample code, so its output is a set of principles applied to your situation. A useful test is to give the agent a real artefact from your team, such as a pull request description or a test plan, and ask it to review that against the skill's focus areas.

Expect the agent to ask about your system, your rollout process and your existing tests, because the skill itself does not supply those details.

• Ask for a review of a change that focuses on what could break in production, not on formatting.
• Ask for a list of edge cases a new interface needs assertions for.
• Ask for acceptance criteria rewritten as measurable conditions.

What does the skill need, and what are the safety notes?

The skill lists no requirements. It does not run scripts and does not use the network, according to its listing. Because it is plain text guidance, there is no executable code to review beyond the SKILL.md file itself.

Check the file before you rely on it. The source is in Japanese, so a team that works in English may want to read it or ask the agent to summarise it first.

• Requirements: none listed.
• Scripts: none run.
• Network: none used.
• Licence: MIT.

What tips help you get more from it?

Use the skill for team-level decisions, such as how to review AI-generated changes, not for single-line fixes. Give the agent your real context, including your test suite, deployment process and interface definitions, so its advice refers to something concrete.

Treat the output as a checklist to discuss with your team, not as a fixed standard.

• Pair it with your own review and testing rules, since it sets principles rather than tooling.
• Ask for the review focus list to be applied to one recent change to see how it reads in practice.
• If you work alone on small scripts, skip it, because the process model adds overhead without much benefit.