Skip to main content

Skill Reference Standard

The playbook is the reference layer for building skills.

A skill should carry doctrine the playbook has already made clear: the capability, human edge, AI delegation boundary, proof signal, and receipt. The public page does not need to point into agent internals. The operational skill points back to the public reference.

Core Rule

Skills may reference playbook pages. Playbook pages should not depend on private skill paths.

That direction matters. The playbook stays readable and public. Skills stay operational and testable.

The learning return path carries proven reusable teaching, not operational instructions:

Playbook doctrine -> operational skill -> measured use/eval
-> proven reusable lesson -> Playbook enhancement -> reindex

Stackmates owns the skill, its evals, controls, and private receipts. The Playbook owns the public explanation. A skill may reveal that its referenced page is unclear, incomplete, or contradicted by evidence; that is a proposal to improve the public owner, not permission to publish the skill body.

Skill Reference Contract

Before a skill is written or materially improved, name:

FieldQuestion
Playbook referenceWhich public page is the source doctrine?
Capability or doctrineWhat capability, standard, or loop is being operationalized?
Human edgeWhat judgment, trust, taste, responsibility, or relationship remains human?
AI delegation boundaryIs the agent doing, assisting, orchestrating, automating, or governing?
Skill outputWhat artifact, decision, action, or review does the skill produce?
Proof signalWhat observable evidence shows the skill worked?
ReceiptWhat should the next loop inherit?

Use the delegation ladder:

Do -> Assist -> Orchestrate -> Automate -> Govern

How To Use

When building a skill:

  1. Start from the closest playbook page.
  2. Extract the capability, human edge, and proof signal.
  3. Convert the doctrine into a skill output contract.
  4. Add evals or gates that can fail when the skill misses the proof signal.
  5. Emit or name the receipt the next loop should reuse.

When a measured run reveals a reusable doctrine gap:

  1. Separate the public lesson from repository-specific procedure and private evidence.
  2. Confirm the lesson is proven, transferable, and belongs to the referenced page rather than a new duplicate.
  3. Route the smallest teaching improvement through the Playbook editing capability.
  4. Re-run the skill eval that exposed the gap and the publication checks for the changed Playbook surface.
  5. Reindex public projections only after the source page is correct.

If no public reference exists, improve the playbook before treating the skill as canonical.

Failure Modes

  • The skill invents doctrine that no playbook page explains.
  • The playbook links into private skill paths and becomes an operator manual.
  • The skill has a trigger but no proof signal.
  • The skill outputs text but no receipt, gate, or reusable artifact.
  • The agent automates work without naming the human edge it must preserve.
  • A measured skill lesson is copied into the Playbook with private paths, controls, or client evidence intact.
  • The Playbook is edited from an unproven observation or reindexed before its source page is correct.

Context

Questions

Which playbook page would a future agent read to understand why this skill exists?

  • What proof signal would show the skill improved the loop?
  • What human edge would be damaged if the skill automated too much?
  • What receipt lets the next skill or flow start from proof instead of memory?