Human approved publishing

Human-approved publishing

Use this skill when an agent may prepare material for publication but must not perform the external action without a clearly recorded human decision.

Inputs

• Draft content or structured material.
• Target channel and audience.
• Publication requirements: citations, formatting, validation, privacy, and voice.
• The identity of the human approver or approval mechanism.

Outputs

• A publication candidate.
• A validation report with blocking and non-blocking findings.
• A decision record with one state: draft, ready_for_review, approved, changes_requested, or published.
• If approved, an exact action packet describing what would be published and where.

See examples/decision-record.json.

Workflow

1. Confirm the target channel, intended audience, and irreversible action.
2. Create or normalize the draft without publishing it.
3. Run deterministic checks available for the artifact.
4. Review facts, privacy, attribution, links, tone, and channel-specific constraints.
5. Present:
   • the exact candidate;
   • blocking findings;
   • non-blocking suggestions;
   • the exact external action awaiting approval.
6. Stop in ready_for_review.
7. Move to approved only after an explicit human approval tied to the current candidate.
8. Before execution, verify that the candidate has not changed since approval.
9. Record the result as published only after the external system confirms success.

Approval validity

Approval is scoped to the exact artifact, destination, and action presented. A material content change, different destination, or new external side effect invalidates the prior approval and returns the state to ready_for_review.

Failure handling

• If validation fails, use changes_requested and list the blockers.
• If the approval is ambiguous or refers to an older version, remain in ready_for_review.
• If publication fails, preserve approved, record the failure, and do not claim published.
• If external state cannot be verified, report publication_unconfirmed.

Boundaries

• Never infer approval from silence, prior enthusiasm, or approval of a different draft.
• Never hide the target destination or external side effect.
• Never label a prepared artifact as published.
• Do not weaken deterministic validation to obtain a green result.
• Do not include secrets or private verification evidence in the public artifact.