kdna-creator git:20260914.a5ebaa1A

Immutable. This exact content never changes and is served at /api/v1/blob/630624d1455fa892.

---
name: kdna-creator
description: Help an Agent author ordinary or explicitly structured judgments from authorized text/interview materials through the current Studio CLI, with actual human or expressly delegated Agent adoption and saved-byte completion.
---

# KDNA Creator

Use the current approved Studio CLI as the creation authority. This skill guides substantive authoring; Core owns component interpretation and Studio owns materialization and the one-use lifecycle. Read [the terminal adapter](references/terminal-session.md) before starting. Exact archive content and a Host-approved installation are required; no global binary, latest package, alternate checkout or old-Core fallback.

Start from the user's natural language and reuse the existing authorization. Establish the intended judgment, scope, explicit materials and new private output directory. Ask only for a material ambiguity or genuinely unsupported boundary. Do not ask people to author protocol IDs, carriers, hash fields or review tickets. No material path means no directory scan. Read only named UTF-8 text files or interview through the selected live channel. If verified-local processing is required and the Host cannot attest it, stop before reading; a local executable does not prove where the Agent processes content.

Draft at least two substantive alternatives for each intended judgment. Explain evidence, applicability, uncertainty and tradeoffs. Preserve the actual source and license/notice obligations. New comparisons are new editorial work, not retrospectively proven producer loss. Do not manufacture extra judgments, taxonomy, diagnostic structure or prerequisites to fill a template. A method may be absent. Use one of the frozen supported component forms only when it expresses the actual judgment. Keep candidate-specific conditions with that candidate, never AND them into overall conditions; an explicitly empty overall condition set may be correct.

Obtain actual selection and fresh final adoption of the full current preview. Use a separate human channel for human-declared review. When the user or authorized supervisor has expressly delegated editorial adoption to an Agent, use the separate Agent adoption channel and record that existing authority. Represent the actor as an Agent, never as a human. A synthetic callback executing is a test, not an actual editorial decision. Do not invent a missing reply or replay an old content decision into a new context.

The Agent interprets actual reviewer language; role, message/review coordinates and final context are bound by the live Host. Neither channel proves identity. Keep stdin/stdout private because it contains material bodies, alternatives, replies and evidence. After any revision, select current alternatives again, show a new preview and obtain fresh final adoption. EOF/error/abort terminates the workspace; do not reconstruct a live capability from JSON.

Deliver the exclusive private bundle only after the CLI saves, reads back and completes the actual bytes. A live `accepted_with_live_context` result is limited to that creation and saved-byte check. Static reopening keeps `creation_accepted:not_evaluated`, `live_context:unavailable`, `identity:not_verified` and `action_authorization:not_evaluated`. Actual editorial acceptance, editorial fitness, identity, action authorization and publication are separate evidence. Read still requires explicit permission and grants no action authority.

Current format2 and its exact graph are required. Historical format1/format-absent artifacts keep their original accepted release and UNKNOWN missing pins; never infer from version, relabel evidence, replace an external expected binding or silently import a prior workspace. See [the typed input guide](references/component-creation-draft.md) for the current authoring surface.