Skip to main content
When creating, optimizing, or combining Skills, calling /comet-any in the Agent and confirming prompts is enough in most cases. Evaluation, publishing, distribution, and recovery are guided by the flow. Unless you are troubleshooting or building automation integration, you do not need underlying commands, internal state, or package structure first. This page is advanced content. For a fast path, start with Quick Start: Combining Any Skill.
/comet-any is Comet’s Skill creation entry point (also known as Skill Creator). You describe the desired Workflow that reads the real local Skill, compiles the target into a Workflow Contract, and regenerates into a ** verifiable, reviewable, and distributable ** Skill Bundle And hide the back-end complexity (Bundle, Factory, composition, resolved skills) internally.

Core design principles

The key to understanding /comet-any is a ** two-layer separation ** The only entry point that users need to remember is /comet-any. When it is necessary to restore the creation state, use comet creator; After entering the release, use comet publish; comet bundle is only reserved as an advanced backend for auditing and troubleshooting.

Main path

The expanded sequence is:
comet creator inspects and resumes the /comet-any creation workflow. comet bundle is the advanced Bundle backend. comet skill is a low-level Skill tool, suitable for local debugging and Engine Run. None of them are the main entry points for creating skills on a daily basis.

Three starting points

The user only needs to choose one of the three starting points at the first level. They correspond to three **Skill Creator intents ** (skillCreatorIntent), and the back end compiles each path to the same Workflow Contract:

Workflow Contract: A common model for all paths

No matter which starting point you choose, /comet-any will eventually compile the target into the same Workflow Contract. This is the core concept for understanding the entire Skill Creator:

Xiaoyu organized the three starting points of new creation, upgrade and combination into the same Workflow Contract template

Whether you start by creating, upgrading, or composing, the result is the same Workflow Contract.

Two types of Workflows

comet-five-phase-overlay

Make incremental adjustments based on the classic five-stage framework of Comet. Retain the main process of /comet and the state semantics of .comet.yaml, with 8 built-in nodes.

workflow-kernel

Brand-new custom workflow. When declaring nodes, Output Schema, Guardrail and Handoff from zero, the state needs to be redeclared.

Protected boundary: Five-stage customization based on Comet’s existing Skill

When you choose “Five-stage Customization based on Comet’s existing Skill”, /comet is regarded as the “protected boundary”. comet-five-phase-overlay has these 8 built-in nodes: There are clear rules in the normal mode:
Want to replace the control node (open/execute/verify/archive)? Normal mode does not allow it. At this point, you need to switch to the advanced workflow-kernel and re-declare the state, Output Schema, and Guardrail.

8 built-in Output schemas

/comet-any has 8 built-in Output schemas. runtime/eval/readiness determines whether a node meets the standards based on the Schema. Skill names are only used for identification.
When a producer node performs override, it must declare in satisfies which Output Schema it satisfies. For example, to replace the design implementation, you must satisfies: [“comet.design.v1”], otherwise the verification will report producer-missing-output-schema.

Solution example: Add Gray-Me requirement clarification to /comet

The following plan.json uses Required Skill Call to force the invocation of grill-me on the open node, allowing /comet to clarify the requirements before entering the design - ** without replacing any node implementation ** :
Key point: The implementation Skill of open remains unchanged. It only forces the invocation of the additional Skill within the node and retrievates the clarification evidence through the Output Schema.

Custom Node (workflow-kernel

workflow-kernel allows the declaration of brand new nodes. Each node must define its responsibilities with responsibility and declare the Output Schema and Guardrail:

What does comet-any do

A complete call to /comet-any will do these things in sequence:
  1. First, try to restore the current creative state; Create only when there is no recoverable state.
  2. Read the preference order in .comet/skill-preferences.yaml (advisory/strict mode).
  3. Use find-skill ** to parse the real Skill content **, including SKILL.md, references, rules, scripts and hooks; The name is only used to locate the candidates.
  4. When dealing with ** missing or ambiguous ** candidates, it is necessary to suspend asking the user.
  5. Compile the target into a Workflow Contract (Workflow Nodes, Skill Bindings, Output Schemas, Guardrails, Handoffs).
  6. ** Verify ** Workflow Contract (control cannot override, producer override must satisfy, Output Schema must exist, etc.).
  7. First, display the **Skill Creator Solution Confirmation Page **, and write the draft only after the user’s confirmation.
  8. Generate entry skills, internal skills of each node, reference/workflow-protocol.json, scripts and platform agent definitions through authoring lane.
  9. Generate the comet/eval.yaml evaluation list (including the Engine Package when Engine is enabled).
  10. Check readiness through comet creator status/next, then proceed to review, approval, publish and distribute.

When to use /comet-any

  • I want to turn the team process into a reusable Skill.
  • We want to optimize the existing skills to make them evaluable and releasable.
  • Want to add, replace or close certain links on the basis of the five stages of /comet.
  • Want to combine multiple skills and ** retain genuine source evidence **.
  • I want to distribute the products to platforms such as Claude Code and Codex.

/comet-any output

After a complete generation or optimization, the product includes:

Responsibilities of key documents

workflow-protocol.json is the core. runtime, eval, review, and publish readiness all read this protocol to ensure that “validation at creation time” and “recovery at runtime” point to the same fact.

required capability set

The required capability set of the stable combination Skill Bundle is skills/scripts/rules/hooks/references/agents:
  • skills: Entry Skill and internal Skill of each node.
  • scripts/rules/hooks: required control plane, not an appendage file that can be deleted at will. They handle state advancement, guards, and handovers deterministically.
  • references: Genuine source evidence and creation audit.
  • agents: Native agent definition of the platform. Claude Code distribution writes the generated custom agents to the target platform preview.
hooks/*.yaml is the Comet portable hook descriptor, and it will only take effect after being compiled to the target platform configuration through comet publish distribute.
/comet-any should not be used as the user’s main process for manual comet bundle commands. Bundle is a deterministic backend for internal use that /comet-any calls automatically.

hard gates

/comet-any has several uncompromising hard rules. Understanding them can help you avoid falling into traps:
  • The user only invokes this Skill, and the CLI is the internal backend.
  • It is necessary to ** parse the real Skill with find-skill **, and it is not allowed to infer the ability merely by name.
  • Missing or ambiguous candidates must be paused. Ask the user and never silently ignore or select for the user.
  • The confirmation page of the proposal must be displayed first. Only after the user’s confirmation can the draft be written.
  • Required Skill Call ** does not replace ** node implementations.
  • producer override must declare the Output Schema of satisfies.
  • The normal mode of the control node cannot override.
  • eval, review, and publish readiness must read ** the same workflow-protocol.json.
  • The products must not retain the AUTHORING PENDING mark. If the entry Decision Core or substance node is not completed, the release will be blocked.
  • The sub-agent Handoff must require the sub-agent to load the Required Skill Call and return the evidence.
  • The script only reads the protocol and state, and does not take the Skill name as the verification basis.
  • “Eval skipped or failed → Never enter ready.”
  • “Lack of manual approval → Never ready.”
  • “preview is necessary before distribution. Only after user confirmation can it be written.”
  • Before installation, the user must be consulted. Automatic installation is not allowed.

validation findings code

/comet-any will verify when compiling the Workflow Contract. If it fails, a clear finding code will be given:

Next step

  • The complete process of creating a Skill ](/en/skill-creator/workflow) - from preparing preferences to generation, evaluation, release, and distribution
  • Skill Preferences and Real sources - Configuration .comet/skill-preferences.yaml
  • publish and distribute Skill](/en/skill-creator/publishing) - readiness, approval, publish and distribute
  • Skill and Engine](/en/skill-creator/engine) - Skill package structure and Engine runtime
  • Overview of the Evaluation System ](/en/eval/overview)-Eval: How Does eval Provide Evidence for Release
Last modified on September 4, 2026