01
Load Project Memory
Read .kiyo/project/ and recorded working rules before touching source code.
File-based · Markdown · No runtime engine
Kiyocontrols AI coding agentsthrough Markdown.
Project Memory first. One mission at a time. Runtime is the judge. A control layer for Claude Code, OpenAI Codex, and GitHub Copilot built on plain files — no runtime service, no database, no autonomous agents.
New to Kiyo The Learn page explains everything in plain English — beginner-friendly, with a technical section for developers.
Claude Code · OpenAI Codex · GitHub Copilot No database No app dependency Portable across tools
15
Manual mission skills
15 + 1
Auto + reference skills
8
Specialist Agents
0
Runtime engines
Framework map
How every command and skill connects.
Teal nodes are commands you invoke with /kiyo-<name>. Blue nodes are discipline skills the host applies automatically. Amber nodes are stack skills. Drag a node, scroll to zoom, or click one to see exactly what it hands off to.
Mandatory Priority Order
Project Knowledge Base > Existing Project Pattern > Global Kiyo Skill > Generic Best Practice
Project facts always beat global defaults. If evidence is missing, agents must say "Unknown" — not invent.
AI Workflow
One request becomes one scoped, verified mission.
Kiyo does not run an agent service. It gives the host AI a deterministic sequence of Markdown contracts to read, so the tool can classify, constrain, execute, and report work in a repeatable way.
02
Route one mission
Command Router and Mission Router bind the request to one workflow and one mission type.
03
Resolve safety
Safety Mode and Response Mode are selected from explicit user intent, project facts, command defaults, and risk triggers.
04
Plan scoped execution
Execution Planner records In Scope, Out of Scope, approval requirements, and verification method.
05
Verify or say so
Runtime evidence decides the final Verification Status. No invented command output.
.kiyo/project/Project Memory
Detected project facts and human-corrected knowledge.
~/.kiyo/config/Framework config
Markdown source of truth for safety behavior.
~/.kiyo/workflows/Workflow contracts
Mission intake, execution, and verification steps.
~/.kiyo/agents/Specialist agents
Portable role contracts for focused AI behavior.
plugins/*Host wrappers
Claude Code, Codex, and Copilot entry points.
The real problem
AI can write code. Kiyo helps it write code that fits your project.
The real problem is not code generation. It is project understanding.
Junior developers and AI coding tools often struggle with the same thing — understanding how the real project actually works. They may not know where validation belongs, how services are structured, which database pattern is used, or whether the project actually uses Clean Architecture, CQRS, EF Core, SQL Server, or Angular standalone components. That is how AI-generated code becomes technically correct but wrong for your project.
Before Kiyo
- AI guesses your architecture.
- Junior developers copy patterns from the wrong file.
- Bug fixes turn into accidental refactors.
- Risky database or auth changes happen without enough caution.
- "Done" often means "looks correct" — not verified.
With Kiyo
- AI reads Project Memory first.
- Existing project patterns come before generic best practice.
- Every task starts with Mission Scope.
- Risky changes trigger strict Safety Mode.
- Final responses carry Verification Status and risks.
For junior developers in real codebases
Kiyo gives juniors and AI coding tools the same artifacts a senior would reach for first — Project Memory, existing pattern guidance, Mission Scope, Safety Modes, Runtime Verification, Specialist Agents, and backend / frontend checklists.
Core Principles
Four disciplines that gate every task.
Each principle is a Markdown rule the agent reads and follows. No Confidence Labels — the meaning lives in Mission Type, Safety Mode, and Verification Status.
Project Knowledge Base first
Every agent reads `.kiyo/project/` before changing code. Detected facts beat global defaults — always.
Mission-driven Scope
One mission at a time. Broad requests get narrowed. Multi-mission requests get split. Unrelated work is deferred.
Safety Modes
standard, strict, or deep — resolved per task. deep is read-only by default. Risk triggers escalate to strict.
Runtime Verification
Runtime is the judge. No "fixed" or "tests pass" without evidence. Status is Verified, Implemented but not verified, Not verified, or Could not verify.
Response Modes
Three modes for response length. Safety always wins.
Response Modes pick how long the response is. compact for short low-risk work. standard for normal development. full-evidence for risky, broad, or evidence-heavy work. A Safety Fallback rule overrides compact whenever a risk trigger fires.
compact
compact
Short, low-risk, prompt-only, code-only, summary-only, or follow-up responses. Keeps Verification Status, Approval Required, and risk fields.
standard
standard
Default for normal development. Includes Mission Scope, Safety Mode, Files Changed, and Runtime Verification.
full-evidence
full-evidence
Architecture, audit, migration, security, high-risk, or evidence-heavy work. Includes Evidence Reviewed, Assumptions, Findings, Risks, Approvals, and Uncertainty.
Compact mode saves tokens by reducing wording, not by removing safety.
Response Format
Every coding response carries Mission Scope, Safety Mode, and Runtime Verification.
No invented commands. No fake test counts. The response template makes it impossible to claim "done" without recording evidence.
- Verifieda runtime command was executed and succeeded.
- Implemented but not verifiedcode changed, no runtime command was attempted.
- Not verifieda command ran but the result is inconclusive.
- Could not verifyenvironment, mode, or approval gate prevented verification.
agent-response.md
## Mission Scope
Mission: Fix the off-by-one in OrderService.CalculateTax
Mission Type: Backend bug fix
In Scope: src/Orders/OrderService.cs, tests/Orders/OrderServiceTests.cs
Out of Scope: src/Notifications/**
Approval Required: No
Workflow Used: bug-fix.md
Verification Method: Reproduction or regression test
## Safety Mode
Mode: standard
Reason:
- Backend bug fix mission, no risk triggers detected.
Escalation:
- None.
## Response Mode
Mode: standard
Safety Fallback: none
## Runtime Verification
Verification Status: Verified
Commands Run:
dotnet test --filter Orders
Results:
| Command | Outcome | Notes |
|--------------------------|---------|--------------------|
| dotnet test --filter ... | pass | 12 passed, 0 fail |Commands
15 mission commands. Read-only by default where it matters.
Invoke as /kiyo- commands on Claude Code, $kiyo- commands on OpenAI Codex, or /kiyo- commands on GitHub Copilot.
/kiyo-initstrictInitialize the current repository for Kiyo. Generates Project Memory and onboarding guidance.
/kiyo-statusdeepRead-only readiness check. Reports a Readiness Level and Recommended Next Action.
/kiyo-change-impact-analyzerdeepAnalyze change impact, risks, affected layers, implementation order, and AI-ready tasks before implementation.
/kiyo-handover-document-generatordeepGenerate a read-only technical handover guide for maintainers, support, testers, and AI agents.
/kiyo-bug-fixstandardLayer-detecting bug fix with cited root cause evidence and regression coverage.
/kiyo-refactor-surgeonstrictFocused, behavior-preserving refactor within an approved scope.
/kiyo-new-requirementstandardImplement a new requirement across detected UI, API, backend, test, and data-flow patterns.
/kiyo-test-writerstandardWrite or update tests in the project's existing testing style.
/kiyo-architect-plannerdeepPlan architecture options, boundaries, trade-offs, and implementation sequence.
/kiyo-audit-specialistdeepRead-only audit for code, configuration, architecture, maintainability, or risk.
/kiyo-legacy-specialistdeepInvestigate legacy or unclear code and recommend incremental modernization.
/kiyo-migration-specialiststrictPlan or review risky migrations with verification and rollback discipline.
/kiyo-performance-investigatordeepRead-only performance investigation with measurement discipline.
/kiyo-security-reviewerdeepRead-only security review with severity, threat model, and source-to-sink tracing.
/kiyo-dependency-security-auditordeepRead-only third-party dependency security audit for .NET and Node manifests.
How It Works
Three layers. Same agent. Different tools.
01
Specialist Agent
A Markdown role contract under the framework reference agents folder. Portable across Claude Code and OpenAI Codex.
02
Role Template
Reusable prompt blueprints under the framework templates when an agent needs one. Stack-agnostic.
03
Host Wrapper
A skill wrapper per host — /kiyo- slash commands on Claude Code, $kiyo- skills on OpenAI Codex, /kiyo- prompt files on GitHub Copilot.
Try Kiyo in your repository.
Install the Claude Code plugin, the OpenAI Codex package, or the GitHub Copilot package, open your target project, run kiyo-init, review Project Memory, then check readiness with kiyo-status.
Claude Code
/plugin marketplace add ./
/plugin install kiyo@kiyo-framework
# restart Claude Code
# open target project
/kiyo-init
/kiyo-statusOpenAI Codex
cp -r plugins-codex/kiyo/skills/* ~/.agents/skills/
cp plugins-codex/kiyo/AGENTS.md ./AGENTS.md
# open target project
$kiyo-init
$kiyo-statusGitHub Copilot
cp plugins-copilot/kiyo/copilot-instructions.md .github/copilot-instructions.md
cp -r plugins-copilot/kiyo/prompts/* .github/prompts/
cp -r plugins-copilot/kiyo/instructions/* .github/instructions/
# open target project
/kiyo-init
/kiyo-statusOpen call
Share your Kiyo skill with the community.
Have a stack-specific guide, a domain discipline, or a new auto-trigger pattern Send it to us and we'll consider it for the next release. We want the skill library to grow with real use cases — not just our defaults.