Specification

Spec version 0.1 · reference runtime agentguides 0.5.10

This page defines the Guide format: the files you author, the data model, the JSON schemas that validate them, and how validation works. To install and run the runtime, see the Quickstart.

  1. What a Guide is
  2. The data model
  3. Schemas
  4. Validation
  5. Full specification

What a Guide is

A Guide is a Skill (per the Agent Skills spec) whose metadata.type is guide. On top of the Skill, it adds a workflow definition, an ordered set of steps, and a runtime record of each walk:

my-guide/
  SKILL.md        # Skill-spec compliant; metadata.type: guide
  GUIDE.md        # the workflow: goal, prerequisites, end states, rollback
  steps/
    010-prep.md   # one markdown file per step
    020-execute.md
    030-verify.md
  scripts/        # optional, per the Skill spec
  references/     # optional
  assets/         # optional

Every Guide is a valid Skill. A Skills-only harness reads it as an ordinary Skill; a Guide-aware harness executes the workflow step by step. Because a Guide is data rather than framework code, the same Guide runs across clients.


The data model

Artifact Authored? What it carries
SKILL.md (+ guide extension) yes Skill frontmatter; metadata.type: guide and the metadata.guide block
GUIDE.md yes id, version, summary, goal_state, prerequisites, end_states, rollback_strategy
steps/NNN[a-z]?-<id>.md yes one step each: performer, action, verification, failure handling, interactions
BOOK.md yes (optional) groups related Guides into a Book; declares child Guides and their roles
run state file no — written by the runtime one run’s frontmatter + the append-only event log

A few concepts the rest of the spec leans on:

  • performer — every step declares who acts: agent, human, or either. A single Guide mixes them freely. Neither party is the other’s safety net; both are co-protagonists.
  • verification — how a step is judged done: script, agent_judgment, human_confirm, or none. An agent_judgment step must record its reasoning in the log before declaring a verdict.
  • end states — a Guide declares its possible exits, each with an optional score (the Postgres example scores migrated at 1.0, a clean rollback to original at 0.5). A rollback is still an end state — failure is recorded, not hidden.
  • failure handling — a step’s on_failure can retry, recover (hand off to a rescue Guide, then optionally resume), ask, or abort.
  • run / walk / event log — a run is one execution, persisted as a single run file with a timestamped, append-only event log. To walk a Guide is to traverse it; the run file is the evidence that traversal leaves behind.

Schemas

The normative contracts are JSON Schema (Draft 2020-12), generated from the runtime’s models and served here so the $id URLs resolve. Editors that honor the # yaml-language-server: $schema=... directive validate your frontmatter as you type.

Authored artifacts — the files you write:

Schema Validates
guide.schema.json GUIDE.md frontmatter
step.schema.json a step file’s frontmatter
skill-guide-extension.schema.json the metadata.guide block in SKILL.md
book.schema.json BOOK.md frontmatter

Runtime + derived shapes — produced by the runtime, not hand-written:

Schema Describes
run-state.schema.json a run file’s frontmatter
walk-skill.schema.json the rendered walk Skill’s guide.runtime block
walk-path.schema.json a reconstructed traversal of one run
dag.schema.json the step dependency graph

Schema URLs are versioned by spec major.minor (/schemas/0.1/). A minor or major bump mints a new path segment; patch-level edits (wording only) overwrite in place, since the set of valid documents is unchanged.


Validation

guide validate --root <path> checks a Guide (or Book) against the schemas above, verifies the step dependency graph is a DAG, and resolves every reference (requires, guide_ref, recover_with). It auto-detects whether the target is a Guide or a Book. See the Quickstart for the command in context.


Full specification

The complete normative specification — exact frontmatter fields, the step lifecycle, verification and recovery rules, and the runtime contracts — is maintained in SPEC.md. It covers:

  • Compatibility with Agent Skills
  • SKILL.md extensions · GUIDE.md schema · step file schema
  • Step lifecycle · verification · failure recovery · rollback
  • Interactions and free-form result reporting
  • State and audit model · the StateBackend interface
  • Composition · reference resolution · versioning
  • The harness contract (what a new client must implement)
  • Appendix — a worked example (Postgres major-version upgrade)

© 2026 Brian Cripe, Agent Guides
MIT-licensed · GitHub