KDNA开源AI工作流

→ Codex: detected, kdna-loader installed

→ Claude Code: detected, kdna-loader installed

→ OpenCode: not detected (install opencode first)

→ Installed domains: 1

```

Want to create your own? kdna init my_expertise scaffolds a minimal domain. Then kdna validate my_expertise checks it, and kdna publish my_expertise sends it to the registry.

---

30-second example

User: "Help me improve this product launch post."

Without KDNA:
  → Suggests clearer wording, shorter sentences, more enthusiasm.

With writing.kdna:
  → Classifies as structural writing diagnosis (not language polishing).
  → Checks: Is there a real argument? A cognitive hook? Evidence density?
  → Avoids banned terms: "polish the language", "make it punchy".
  → Self-checks: 5/5 passed. Risk flags: none.

The agent didn't get better at writing. It got better at judging what kind of problem this is.

---

5-Minute Quick Start

One path. Five minutes. See KDNA change an agent's judgment.

npm install -g @aikdna/kdna-cli
kdna setup
kdna install @aikdna/writing
kdna verify @aikdna/writing --judgment
kdna compare @aikdna/writing --input "help me improve this post"

That's it. Your agent now loads a domain judgment package. The last command sends the same input to an LLM with and without KDNA, diffing the judgment paths so you can see exactly what changed.

```bash

FAQ

<details> <summary>Is this related to biological kDNA, KnowledgeDNA, or other DNA projects?</summary>

No. In this project, KDNA refers to the KDNA Protocol: an open judgment protocol for AI systems. It focuses on human-governed domain judgment assets, not biological kinetoplast DNA, goal-tracking SaaS, codebase summaries, agent identity profiles, or model lineage analysis. </details>

<details> <summary>How is KDNA different from Prompt, RAG, Skills, and MCP?</summary>

• Prompt changes what AI says — task-scoped, ephemeral.
• RAG changes what AI can access — retrieves facts.
• Tools/Skills/MCP change what AI can do — connect to actions.
• KDNA changes how AI judges — domain-scoped, version-controlled, auditable.

KDNA is the judgment reference layer. It does not replace these mechanisms — it sits alongside them. </details>

<details> <summary>Is KDNA just a fancy system prompt?</summary>

No. System prompts are free-text behavioral instructions scoped to a single conversation. KDNA is a structured, validated, version-controlled format with explicit fields (axioms, boundaries, self-checks, failure risks). KDNA packages are designed to be signed, hash-verified, and distributed through a registry — a system prompt is none of these. </details>

<details> <summary>Does KDNA replace the model?</summary>

No. KDNA is a judgment reference that the model loads before it reasons and acts. The model still does all the reasoning, generation, and tool use. KDNA tells the model what to pay attention to, what to avoid, and what to verify — it does not generate output. </details>

<details> <summary>How is KDNA different from RAG?</summary>

RAG retrieves facts and documents for the model to reference. KDNA encodes what matters and what to watch for in a domain. RAG says "here's the coding standard document." KDNA says "when reviewing code, classify whether the problem is structural or cosmetic before suggesting changes." </details>

<details> <summary>Can I use KDNA without coding?</summary>

Yes — start with KDNA Studio (@aikdna/kdna-studio) for guided domain authoring. To install and use domains with your AI agent, you only need the CLI (npm install -g @aikdna/kdna-cli). Creating your own domain currently requires editing JSON files, though KDNA Studio's interview mode helps non-technical experts structure their judgment without writing code. </details>

<details> <summary>Does KDNA work with any AI model?</summary>

KDNA is model-agnostic. The format encodes judgment as structured JSON — any agent framework that can load context before reasoning can use KDNA. Currently supported agents include Claude Code, Codex, OpenCode, Cursor, and GitHub Copilot. </details>

<details> <summary>What happens if I load multiple KDNA domains that conflict?</summary>

KDNA's composition mechanism detects and reports conflicts rather than silently merging incompatible principles. When domains conflict — for example, a brand domain encouraging emotional intensity while a compliance domain requires conservative wording — the agent reports the conflict rather than choosing one side. </details>

<details> <summary>Can AI agents modify KDNA domains?</summary>

No — not the judgment-class fields. KDNA's Human Judgment Lock requires human confirmation before axioms, boundaries, risk models, failure criteria, and other judgment-class fields can be modified. Operational fields like usage statistics can be updated automatically. </details>

<details> <summary>Where can I see KDNA in action?</summary>

Run kdna compare @aikdna/writing --input "help me improve this post" to see a side-by-side judgment path comparison. Visit aikdna.com/benchmark for evaluation data across multiple domains. </details>

---