Skill Forge

Skill Forge

Outcome

Design one narrow capability as a portable Agent Skills package with explicit activation boundaries, then compile and prove it when a Foundry runtime is available.

Use this skill when

• A successful workflow should become reusable.
• A user asks to create, improve, or package a skill.
• A repeated task needs stable instructions, tools, guardrails, or evaluation.
• The capability should work in ChatGPT, Codex, or a skills-based plugin.

Do not use it when the work requires persistent decision rights, a distinct durable memory scope, or an ongoing trigger owned by an autonomous actor. Use $agent-forge only after proving one of those boundaries.

Procedure

1. Define the decision contract.

   • Capture the objective, audience, deliverables, stakes, reversibility, permissions, constraints, and completion tests.
   • Ask only for missing information that materially changes the capability.

2. Discover before creating.

   • Inspect the capability graph and existing skills.
   • Reuse or depend on an existing capability when it already owns the behavior.
   • Record required, preferred, and forbidden capabilities explicitly.

3. Bound activation.

   • Write concrete positive triggers and at least one negative trigger.
   • Keep one skill responsible for one coherent outcome.
   • Avoid broad adjectives as requirements; translate them into observable gates.

4. Author the Skill Pack.

   • Give every procedure step a proof obligation.
   • Declare inputs, outputs, dependencies, tool policy, memory policy, deployment targets, and optional taste profile.
   • Keep SKILL.md frontmatter limited to name and description.

5. Select the execution mode.

   • Foundry runtime available: use this mode only when tools/foundry/cli.mjs and foundry/contracts/ exist in the current workspace. Validate the Task Envelope and Skill Pack, then run:

     node tools/foundry/cli.mjs forge \
       --envelope <task-envelope.json> \
       --pack <skill-pack.json> \
       --out <package-directory>
     

   • Portable skills-only install: read references/portable-contracts.md, then author the Task Envelope, Skill Pack, and proposed SKILL.md directly with those field names. Mark compilation and proof pending-runtime; do not invent paths, commands, or receipts. Give the operator the command above as an optional handoff for a later SIS workspace run.

6. Prove.

   • Run this step only in Foundry runtime mode.

   • Run static and artifact checks first.

   • Execute command tests only when the Task Envelope permits the executable.

   • Keep manual and judge tests pending until independent evidence exists.

   • Run:

     node tools/foundry/cli.mjs prove <package-directory>
     

7. Promote only from receipts.

   • validated: every required lane passed.
   • experimental: required evidence is still pending.
   • revise: a required test failed.
   • rejected: a critical security test failed.

Quality gates

• The skill is narrower than an agent and more reusable than a one-off prompt.
• Trigger examples distinguish when to use and when not to use it.
• The procedure is executable without hidden context.
• Claims about research, tests, writes, or deployment have receipts.
• The producer is not the sole judge for a required taste lane.

Return

Return the Task Envelope, Skill Pack, proposed skill, execution mode, and unresolved proof. Include a compiled package path and Evidence Receipt only when those artifacts were actually produced.