Tailor explanation depth with ak:coding-level

Use ak:coding-level to persist how much explanation, teaching structure, and strategic context you want in future responses. The selected level is stored as codingLevel in .claude/.ck.json; it changes communication style, not the project's code standards or safety rules.

Choose ak:coding-level for a durable preference

Use ak:coding-level when

• Explanations are consistently too basic or too terse for you.
• You want beginner teaching, professional design context, or lead-level risk framing to carry into later sessions.
• You want a project-scoped preference instead of repeating response-format instructions in every prompt.

Choose another workflow when

• You need a one-response tone adjustment; state it in the request instead.
• You need to change linting, language, architecture, or review requirements. Coding level does not override repository instructions.
• You need a runtime-independent guarantee. Automatic injection is evidenced by the Engineer session-start Hook, whose full support is Claude-specific.

Prepare the preference

• Complete Onboarding, and confirm Engineer Kit is installed in the intended project scope.
• Choose an explicit integer from 0 through 5.
• Preserve unrelated keys in .claude/.ck.json.
• Start a fresh supported session after changing the value when you want to verify automatic injection.

See Runtime adapters before relying on a Hook-backed behavior outside Claude Code.

Set the level

The Skill accepts one level. Its metadata disables automatic model invocation, so select it explicitly rather than expecting the assistant to change your preference on its own.

/ak:coding-level 2

/ak:coding-level 2

$ak:coding-level 2

No behavior for an omitted or out-of-range argument is specified. Provide a valid level or ask the runtime to clarify instead of guessing.

Understand what happens during a run

1. You select the level. The Skill receives an explicit value from 0 to 5.
2. The preference is persisted. It updates codingLevel in .claude/.ck.json while preserving the rest of the project configuration.
3. A supported session starts. The Engineer session-start Hook loads the config and defaults to level 5 when codingLevel is absent.
4. The Hook maps the level to guidance. It exposes the selected style name and injects the corresponding communication rules into session context.
5. Responses follow the profile. Explanation depth, code-example shape, and response sections change; repository constraints and approval boundaries do not.

Claude Code also exposes the six named manual styles through the Output style setting in /config. Native installs emit them into the runtime root's output-styles/ directory; explicit plugin installs emit them into the plugin root's output-styles/ directory. The session Hook resolves that active layout first and then the legacy .agentkit/output-styles/ sidecar. If neither contains the selected style, it emits a diagnostic instead of silently omitting the guidance.

Current /output-style discovery is not verified by an AgentKit runtime canary. The source does not establish manual style selection on Cursor or Codex, so treat it there as unsupported unless the live runtime shows it.

Keep the change scoped

• Review the .claude/.ck.json diff because the preference persists across sessions in that project.
• Do not replace the whole config file to change one key.
• The Skill itself specifies no provider, network call, subprocess, paid service, or large disk effect beyond the small config edit.
• Manual output-style selection can affect the current response shape without proving that the persisted preference was loaded.
• A same-named style that you created is preserved with a warning during a routine ak update. ak kit init --force may replace only that selected-Kit collision; --fresh performs the separate snapshot-backed reset of the selected install target.

Verify the result

A complete run should give you:

• A valid codingLevel value in .claude/.ck.json.
• No unrelated config changes.
• On a supported fresh session, guidance that visibly matches the selected profile.
• At level 1, for example, explanations that lead with why, identify common pitfalls, and include key takeaways for significant work.

The Skill defines no standalone tests. Verify the persisted JSON, restart the supported runtime, and compare an observable response with the selected level.

Troubleshoot or continue

Know the current limits

• Automatic injection depends on the Engineer session-start Hook; the exact sources provide full evidence for Claude Code, not equivalent Cursor or Codex behavior.
• The six profiles are broad defaults, not a measurement of ability.
• The current stable and beta packages contain identical ak:coding-level, output-style, and relevant session-injection content.