15 manual commands, 15 auto skills, and 1 reference skill.

Reads Project Memory first.

Before suggesting or applying any change, it reads project-profile.md, project-structure.md, and ai-working-rules.md (plus any area-specific pattern file) so project facts drive the answer, not generic training data, then applies the Mandatory Priority Order before proposing anything.

Trigger: Code change, refactor, add, or explain in a Kiyo-initialized repo

Escalates to strict Safety Mode and requires approval planning.

Once a risk trigger fires, it forces strict Safety Mode, requires an explicit Verification Plan and Rollback Plan, and blocks the change from being applied until the user approves it. Read-only investigation is still allowed while approval is pending.

Trigger: Migration, schema, auth, secret, production, delete, remove, or upgrade risk

Requires file and line evidence for non-obvious claims.

Every non-obvious claim must cite a specific file:line or Project Memory entry. Confidence Labels such as [high] or [low] are forbidden; the response ends with an explicit Verification Status instead of a vague confidence gesture.

Trigger: Review, audit, explanation, or investigation

Blocks invented project facts and uses Unknown when evidence is missing.

Forbids inventing files, APIs, schemas, services, or a specific stack — Clean Architecture, CQRS, MediatR, EF Core, Angular, SQL Server, and similar defaults are never assumed. When Project Memory or source code cannot confirm a fact, the response writes Unknown or Not enough evidence rather than guessing.

Trigger: Stack, schema, API, library, or architecture decisions

Reads focused evidence instead of the whole repository.

Reads Project Memory first, then a handful of directly relevant source files, and only expands further when a stated trigger requires it — an unclear root cause, conflicting evidence, or a security/auth impact. Generated folders such as bin/, node_modules/, and dist/ are never scanned by default.

Trigger: Investigation, fix, review, add, refactor, or test

Maintains project-local memory under .kiyo/project/memory/.

Keeps a small, git-tracked MEMORY.md index plus on-demand detail files so decisions with a non-obvious "why", user feedback, and deferred findings survive across sessions. It never mirrors what git log or the codebase can already answer cheaper.

Trigger: Past decisions, deferred findings, or non-obvious session facts

Asks before writing the wrong code.

When a request to implement, fix, or refactor has more than one reasonable interpretation, a missing required input, or an unclear scope boundary, it presents the distinct readings and asks which one the user wants instead of silently picking one.

Trigger: A request has multiple reasonable meanings

Keeps the solution small and avoids speculative abstractions.

Forbids speculative features, abstractions built for a single caller, premature generalization, and configuration for cases that do not exist yet. The standing test: would a senior engineer reviewing this diff call it overcomplicated?

Trigger: New code or design work

Keeps the diff limited to the requested change.

Restricts an edit to the lines the request actually names — no drive-by refactors, reformatting, or comment cleanup on adjacent code. Problems spotted outside the change boundary are recorded as Deferred Findings, not fixed in the same diff.

Trigger: Editing existing code

Follows existing structure, naming, test style, and boundaries.

Treats the project's own structure, naming, and validation style as the reference point instead of a textbook architecture. Blocks introducing CQRS, MediatR, FluentValidation, a new ORM, or a new layer unless the project already uses it or the user explicitly asks.

Trigger: A project already has patterns

Requires runtime evidence before claiming Verified.

Requires an actual build, test run, lint, or safe HTTP call before a response can claim a change works, and forbids faking command output. Every code-changing response ends with exactly one status: Verified, Implemented but not verified, Not verified, or Could not verify.

Trigger: Any code-changing task

Routes design work through the UI/UX design discipline when available.

Delegates color, typography, layout, and component-style decisions to the ui-ux-pro-max skill when it is installed, ranking its recommendation below the project's own design system. If the skill is not installed, it says so plainly and falls back to safe baseline guidance instead of inventing palettes or fonts.

Trigger: UI, UX, visual design, or design system work

Keeps replies compact without removing risks, evidence, or verification.

Chooses a Compactness Mode — standard, compact, or ultra — for the reply's wording, defaulting to compact for ordinary work and forcing standard back on for security audits, migrations, and other completeness-critical tasks. It never trims evidence, risk warnings, approval gates, or the Runtime Verification status to save words.

Trigger: Coding, review, audit, planning, or follow-up responses

Scans the Manual mission-skill catalog before generic freeform work starts.

When the user has not already named a skill or /kiyo-<name> command, it compares the stated goal against all 15 Manual mission skills and announces the single best match with its command and default Safety Mode, or asks the user to choose explicitly when two or more commands fit equally well. It never adopts a Manual skill's mission or Safety Mode commitment without that confirmation.

Trigger: A request describes a goal in the user's own words without already naming a skill or command

Gives .NET architecture guidance only when the project supports it.

When evidence supports Clean Architecture, a modular monolith, CQRS/MediatR, and EF Core, it gives layer dependency rules (Api → Application → Domain), solution folder conventions, and pragmatic guidance for controllers, validators, and tests — conditional on evidence, never assumed by default.

Trigger: Project evidence confirms .NET Clean Architecture or CQRS

Provides Angular guidance while still following project patterns first.

Reference skill bundled from Google's official Angular guide (not Kiyo-authored). Covers components, signals, forms, dependency injection, routing, styling, and testing, scoped to the project's detected Angular version.

Trigger: Angular evidence is present