skill-validation

name: skill-validation description: Defines input/output schema validation for skills. Use when creating or validating skills, composing skill chains, or checking skill contracts. Agents load this skill to understand how to validate skill I/O. version: 1

Skill Validation Protocol

Overview

This skill defines the validation protocol for all other skills. It enables type-safe skill composition, runtime validation, and caching. Load this skill when creating new skills, validating skill contracts, or chaining skills together.

Enhanced Skill Frontmatter Schema

Every skill MUST have YAML frontmatter with these fields:

---
# Required fields
name: string              # Unique skill identifier (kebab-case)
description: string       # Single sentence describing when to use
version: number           # Schema version, default 1

# Optional: Input specification
input_schema:
  required:               # Inputs that MUST be provided
    - name: string        # Parameter name
      type: string | string[] | number | boolean | object
      description: string # What this input represents
  optional:               # Inputs that MAY be provided
    - name: string
      type: string | string[] | number | boolean | object
      description: string
      default: any        # Default value if not provided

# Optional: Output specification
output_schema:
  required:               # Outputs that MUST be produced
    - name: string        # Output field name
      type: string        # Type description
      description: string # What this output represents
  optional:               # Outputs that MAY be produced
    - name: string
      type: string
      description: string

# Optional: Validation checks
pre_checks:               # Validate before execution
  - description: string   # What to validate
    validation: string    # How to validate (instruction or condition)

post_checks:              # Validate after execution
  - description: string   # What to validate
    validation: string    # How to validate

# Optional: Caching
cacheable: boolean        # Whether outputs can be cached (default: false)
cache_ttl_minutes: number # Cache lifetime (default: 0, only if cacheable: true)
---

Type System

Supported Types

Type Coercion Rules

• string accepts any scalar converted to string
• string[] accepts a single string (wrapped in array)
• number accepts numeric strings (parsed)
• boolean accepts "true"/"false" strings (parsed)
• object accepts JSON strings (parsed)

Pre-Execution Validation Protocol

Before using a skill, agents MUST validate inputs:

Step 1: Check Required Inputs Exist

For each input in input_schema.required:
  If input.name not provided:
    FAIL with: "Missing required input: {input.name}"

Step 2: Validate Input Types

For each provided input:
  If input value type does not match declared type:
    If type coercion possible:
      Coerce value to declared type
    Else:
      FAIL with: "Type mismatch for {input.name}: expected {type}, got {actual_type}"

Step 3: Apply Default Values

For each input in input_schema.optional:
  If input.name not provided and default exists:
    Set input.name = default

Step 4: Run Pre-Checks

For each check in pre_checks:
  Evaluate check.validation
  If validation fails:
    FAIL with: "Pre-check failed: {check.description}"

Validation Failure Format

When validation fails, return structured error:

validation_error:
  skill: skill-name
  phase: pre_execution | post_execution
  check: description of what failed
  expected: what was expected
  actual: what was found
  suggestion: how to fix

Post-Execution Validation Protocol

After skill execution, agents MUST validate outputs:

Step 1: Check Required Outputs Exist

For each output in output_schema.required:
  If output.name not in result:
    FAIL with: "Missing required output: {output.name}"

Step 2: Validate Output Types

For each declared output:
  If output present and type does not match:
    WARN: "Output type mismatch for {output.name}"

Step 3: Run Post-Checks

For each check in post_checks:
  Evaluate check.validation
  If validation fails:
    FAIL with: "Post-check failed: {check.description}"

Skill Chaining Protocol

Skills can be composed into chains where outputs feed into inputs.

Chain Definition

chain:
  - skill: skill-a
    inputs: { param: value }
    outputs_as: step1
  - skill: skill-b
    inputs:
      data: $step1.result   # Reference previous output
    outputs_as: step2

Schema Compatibility Checking

Before executing a chain, validate compatibility:

For each step after the first:
  For each input that references a previous output:
    source_output = resolve $reference
    Verify source skill's output_schema includes the field
    Verify types are compatible (or coercible)
    If incompatible:
      FAIL with: "Chain incompatibility: {source_skill}.{output_field} 
                  cannot feed {target_skill}.{input_field}"

Chain Example

# Step 1: Read memory
- skill: brain-memory-kb
  inputs:
    operation: read
    namespace: org.global
    query: auth patterns
  outputs_as: memory_read

# Step 2: Generate prompt using memory
- skill: ai-orchestration-prompt-engineering
  inputs:
    task_description: $memory_read.entities[0].body
    constraints: ["must be verifiable", "structured output"]
  outputs_as: prompt_result

Compatibility Matrix

Skill Caching Protocol

Skills marked cacheable: true can have their outputs cached.

Cache Key Format

{skill_name}:{hash_of_sorted_inputs}

Example:

brain-memory-kb:a1b2c3d4e5f6

Where hash is SHA-256 of canonical JSON representation of inputs.

Cache Storage

Cache entries are stored in session memory:

Location: session/current/cache/
Filename: {skill_name}-{short_hash}.md

Cache Entry Schema

---
entity_name: session.current.cache.{skill_name}.{short_hash}
namespace: session.current
category: cache
status: accepted
tags: [cache, {skill_name}]
created_at: 2026-04-04T10:00:00Z
cache_key: {full_cache_key}
cache_ttl_minutes: {ttl}
expires_at: 2026-04-04T10:30:00Z
---

{serialized output as JSON or YAML}

Cache Lookup Protocol

1. Compute cache_key from skill name and inputs
2. Check session/current/cache/ for matching entry
3. If found and not expired:
   Return cached output
4. If expired:
   Delete cache entry
   Continue to execution
5. If not found:
   Continue to execution
6. After execution (if cacheable):
   Write cache entry with TTL

TTL-Based Invalidation

• Check expires_at before using cached value
• Delete expired entries on access
• Session end clears all session cache entries

When to Use Caching

Validation Examples

Example 1: Valid Skill Frontmatter

---
name: brain-memory-kb
description: Unified ai-brain memory+KB operations. Use mode=memory or mode=kb-query.
version: 1
input_schema:
  required:
    - name: operation
      type: string
      description: Operation type - read, write, update, or delete
    - name: namespace
      type: string
      description: Target namespace (e.g., org.global, projects.myapp)
  optional:
    - name: query
      type: string
      description: Search query for read operations
    - name: content
      type: object
      description: Entity content for write operations
output_schema:
  required:
    - name: status
      type: string
      description: Operation result - success, not_found, or error
  optional:
    - name: entities
      type: object[]
      description: Retrieved or created entities
    - name: index
      type: object
      description: Updated index state
pre_checks:
  - description: Namespace follows naming convention
    validation: namespace matches pattern ^(org|projects|session)\..+$
  - description: Operation is valid
    validation: operation in [read, write, update, delete]
post_checks:
  - description: Status is returned
    validation: status is not empty
  - description: Write operations update index
    validation: if operation in [write, update] then index is updated
cacheable: true
cache_ttl_minutes: 5
---

Example 2: Validation Failure

Input provided:

operation: read
# namespace missing!
query: "auth patterns"

Error output:

validation_error:
  skill: brain-memory-kb
  phase: pre_execution
  check: Required input validation
  expected: namespace (string) must be provided
  actual: namespace was not provided
  suggestion: Add namespace parameter, e.g., namespace: "org.global"

Example 3: Skill Chain with Compatible Schemas

# Chain: Memory → Prompt → Validation
chain:
  - skill: brain-memory-kb
    inputs:
      operation: read
      namespace: org.global
      query: security decisions
    outputs_as: memory
    # Output: { status: "success", entities: [...] }

  - skill: ai-orchestration-prompt-engineering
    inputs:
      task_description: "Review these security decisions"
      context: $memory.entities      # object[] → feeds into context
      output_format: structured
    outputs_as: prompt
    # Output: { prompt: "...", validation_rules: [...] }

  - skill: failure-engineering
    inputs:
      system_description: $prompt.prompt
      failure_scenarios: ["network timeout", "auth failure"]
    outputs_as: resilience
    # Output: { recommendations: [...], risk_assessment: {...} }

Schema compatibility check:

• memory.entities (object[]) → prompt.context (object): Compatible via array access
• prompt.prompt (string) → resilience.system_description (string): Compatible

Creating Validated Skills

When creating a new skill:

1. Start with required frontmatter: name, description, version
2. Define input_schema based on what the skill needs to function
3. Define output_schema based on what the skill produces
4. Add pre_checks for input validation beyond type checking
5. Add post_checks for output quality validation
6. Set cacheable: true only if outputs are deterministic for given inputs
7. Test validation by providing invalid inputs and verifying error messages

Validating Existing Skills

To add validation to an existing skill:

1. Read the skill content thoroughly
2. Identify all inputs mentioned in the protocol
3. Identify all outputs the skill produces
4. Derive pre-checks from stated preconditions
5. Derive post-checks from stated guarantees
6. Add schema to frontmatter without modifying content
7. Verify YAML frontmatter is valid