How to use the Engineer System Change agent skill

What is an agent skill?

An agent skill is a folder containing a SKILL.md file with instructions that an AI agent reads when a request matches the skill's description. The agent then follows those instructions for that task instead of its default approach.

The engineer-system-change skill is a single SKILL.md file. Its description tells the agent when to use it: assessing RFCs, issues, designs, features, refactors, migrations, dependency changes, and proposed fields, events, APIs, modules and services.

What the skill makes the agent do

When installed, the agent treats a proposed change as a hypothesis to test rather than a task to build. It reads the actual codebase, checks whether the problem is real and who uses the change, and prefers the smallest sufficient fix.

The skill asks the agent to separate verified facts, inferences and unknowns, and to label material risk claims as VERIFIED, INFERENCE or UNKNOWN. For every proposed durable field, event or API, it requires a named committed consumer and an absence test: which verified scenario fails if the addition is removed. It also considers solutions in a fixed order, starting with no change and documentation or configuration, and only moving toward new abstractions or subsystems when the smaller options fail.

The skill returns explicit verdicts: STOP, REDUCE, REVISE and others. It also asks for observed evidence and a verification plan before any edits, and for verification after implementation.

• Assess whether the change should exist before designing it
• Name the real consumers of any new field, event, API or service
• Choose the smallest sufficient solution and reject pseudo-requirements
• Ask for evidence in proportion to the risk

Install the skill

Download the ZIP for this skill from this page. Unzip it so that the skill folder, containing SKILL.md, sits in one of these locations.

For a personal install across projects, unzip into ~/.claude/skills/. For a single project, unzip into the project's .claude/skills/ folder.

Alternatively, upload the skill in the Claude apps under Settings → Capabilities → Skills.

• Personal: ~/.claude/skills/
• Single project: .claude/skills/ inside the project
• Claude apps: Settings → Capabilities → Skills, then upload

Trigger the skill with natural requests

You do not need a special command. Describe the proposal in plain language and name the artefact. The skill's description lists the situations it covers, so requests that mention an RFC, a feature proposal, a migration or a new field match it most clearly.

Say whether you want assessment only or implementation. The skill states that when you only ask for an assessment, planning or review, the agent makes no project or external-state changes. If you want edits, say so explicitly.

• Assess this RFC and tell me whether we need the new events table. Do not change anything yet.
• Check whether this proposed API field has a real consumer in our repo.
• Here is the issue for adding a retry queue. Is the problem real, and what is the smallest fix?
• Implement the approved change for the cache key fix, then verify it with a failing test first.

Examples of what the agent will check

Given a proposed new status field, the agent should trace the current workflow, find which named component would read the field, and check whether an existing mechanism already exposes the same information. If no committed consumer exists, the skill directs it to remove or defer the field.

Given a proposed migration, the agent should look for existing capabilities, consider whether configuration or a local fix suffices, and ask for a rehearsal and a rollback plan only where the change is irreversible or high-risk.

What the skill needs and its safety notes

The skill lists no requirements. It does not run scripts and does not use the network, according to its listing. Its instructions are plain text that guide the agent's reasoning and edits in your repository.

The skill does tell the agent to inspect the current target revision and relevant discussion, and not to rely on a stale checkout, an RFC alone or remembered architecture. Because the agent can edit files when implementation is authorised, review any diff it produces and keep commit, push and deploy permissions separate. The skill itself says implementation permission is separate from permission to commit, push, deploy, publish or update issues and pull requests.

• Runs scripts: no
• Uses network: no
• Requires: nothing listed

Tips for better results

Give the agent the real artefact: the RFC text, the issue link or the relevant files. The skill penalises conclusions drawn from an RFC alone, so the more of the live codebase the agent can read, the better the verdict.

Ask for the verdict and the evidence labels together. A STOP or REDUCE verdict is only useful if you can see which claims were verified and which were inferred.

Do not use it for renames, formatting or explaining how a file works. The skill says those tasks add overhead without benefit or fall outside its scope.

• Provide the RFC, issue or diff together with access to the repository
• Ask for VERIFIED, INFERENCE and UNKNOWN labels on risk claims
• Use a separate review process for a finished diff