Prompt Skillset

We forge prompts collaboratively, and we execute them deliberately.

Philosophy

This skillset assumes:

• Human intent is discovered, not known upfront
• Conversation is not state
• Disk state is authoritative
• Execution is irreversible and requires explicit consent

Core Invariants

1. Conversation is not state

The agent must not assume continuity or memory across turns. All durable state lives on disk. The agent may only read or modify state inside an explicit skill invocation.

Conversation is not state. Disk is state. Skills are the only legal state transitions.

2. Single canonical prompt on disk

There may be at most one active prompt artifact at any time.

Canonical path: .prompt-forge/active.yaml

3. Execution is destructive

After a successful prompt-exec:

• The canonical prompt artifact must be deleted
• The system must return to a "no active prompt" state

4. Optional execution receipts

To preserve auditability without violating single-prompt rules:

• prompt-exec may write an execution receipt
• Receipts are immutable and never executable

Receipt location: .prompt-forge/receipts/<timestamp>-<hash>.yaml

Member Skills

Skill	Purpose
prompt-forge	Shape, refine, and stabilize intent into the canonical prompt artifact
prompt-exec	Execute the forged prompt exactly as written

No other prompt-related skills are permitted.