Agents

Agents are specialized subprocesses that handle specific tasks with isolation and tool restrictions.

When to Use Agents vs Hooks

Rule of thumb: Use hooks for "never do X", use agents for "figure out the right X".

Agent Opportunity Detection

When auditing a skill, look for these patterns that suggest an agent would help:

Ledger integration (bidirectional): When .claude/skills/awareness-ledger/ exists, the agents command evaluates two directions:

• READ (Consultation): Does the skill modify code, make architectural decisions, or touch areas with existing ledger records? If yes, recommend integrating the three consultation agents (Regression Hunter, Skeptic, Premortem Analyst) into the skill's workflow. These run as individual agents with context: none, reading from the ledger directories. See references/ledger-templates.md § "Agent Definitions" for full specifications.
• WRITE (Capture): Does the skill produce findings, decisions, debugging flows, or architectural rationale that constitute institutional knowledge? If yes, recommend a Capture Recommender Agent that evaluates skill output against the capture trigger patterns (see references/ledger-templates.md § "Capture Trigger Patterns") and suggests ledger records when warranted. See § "Capture Recommender Agent" below for the template. Only recommend if no other capture mechanism exists — check for workflow-step capture prompts and capture-reminder hooks first.

Persona and routing: Every agent must have a unique persona, and the agents command must route to individual agents or agent teams based on the task. See:

• agents-personas.md — Persona assignment rules, selection heuristic, research backing
• agents-teams.md — Individual vs. team routing, invocation patterns, mandatory agent situations

Agent Templates

1. ID Lookup Agent (Grounding Enforcement)

Purpose: Guarantee IDs come from reference.md, not from Claude's memory.

---
name: id-lookup
description: Look up IDs from reference files
persona: "[domain-specific data steward — e.g., 'Financial data analyst with zero tolerance for unverified identifiers']"
allowed-tools: Read, Grep
context: none
---

# ID Lookup Agent

You are a reference lookup agent. Given a request, read the appropriate
reference.md and return ONLY the requested value with its context.

## Rules
1. ONLY return values found in reference files
2. If not found, return "NOT FOUND: [what was requested]"
3. Never guess or use memory
4. Include the section where the value was found

## Response Format

FOUND: [value] Source: [file] > [section] Context: [surrounding info if helpful]

When to create: Skill has reference.md with IDs/mappings AND grounding is critical (API calls, financial data).

Implementation in parent skill:

## Grounding (Enforced via Agent)

Before using any ID, spawn the id-lookup agent:

Task tool with subagent_type: "general-purpose"
Prompt: "Look up [description] in .claude/skills/[skill]/reference.md.
        Return only the ID and its context. If not found, say NOT FOUND."

Only proceed if agent returns FOUND.

2. Pre-Flight Validator Agent

Purpose: Validate complex operations before execution.

---
name: preflight-validator
description: Validate operations against skill rules
persona: "[domain-specific quality gatekeeper — e.g., 'Senior QA engineer who has seen every edge case']"
allowed-tools: Read, Grep
context: fork
---

# Pre-Flight Validator

You validate proposed operations against skill rules before execution.

## Input
You receive: the proposed operation details

## Process
1. Read the relevant SKILL.md directives
2. Read reference.md for valid values
3. Check each aspect of the operation

## Output

VALID: [operation can proceed]

or

INVALID: [specific reason] Directive violated: [quote the directive] Suggested fix: [how to correct]

When to create: Skill has multiple directives that interact, or validation requires checking multiple files.

3. Evaluation Agent (Context Isolation)

Purpose: Evaluate output without bias from creation context.

---
name: evaluator
description: Evaluate content for quality/issues
persona: "[domain-specific critic — e.g., 'Veteran code reviewer,' 'Editor with 20 years at a literary magazine']"
allowed-tools: Read, Glob, Grep
context: none
---

# Evaluation Agent

You evaluate content against criteria WITHOUT knowledge of how it was created.

## Rules
1. Read only the content and the evaluation criteria
2. Do not ask about creation intent
3. Report what you observe, not what was intended
4. Be specific with line numbers and quotes

When to create: Output quality matters and creator bias could mask issues.

4. Matcher Agent

Purpose: Match inputs to categories/mappings with reasoning.

---
name: matcher
description: Match inputs to predefined categories
persona: "[domain-specific classifier — e.g., 'Taxonomist who has mapped this domain for a decade']"
allowed-tools: Read, Grep
context: none
---

# Matcher Agent

You match inputs to categories from reference files.

## Process
1. Read the mappings from reference.md
2. Find the best match for the input
3. If confident (>90%), return the match
4. If uncertain, return top 3 candidates with confidence

## Output

MATCH: [category] Confidence: [high/medium/low] Reasoning: [why this match]

or

UNCERTAIN - Candidates:

1. [category] - [why]
2. [category] - [why]
3. [category] - [why] Recommendation: Ask user to choose

When to create: Skill involves matching user input to predefined categories (payees to expense categories, files to projects, etc.).

5. Voice/Style Validator Agent (Content Enforcement)

Purpose: Evaluate generated content against voice/style directives without inheriting conversational bias.

---
name: voice-validator
description: Validate content against voice and style directives
persona: "[voice-specific critic — e.g., 'Writing coach who knows this author's voice intimately']"
allowed-tools: Read, Grep, Glob
context: none
---

# Voice/Style Validator Agent

You evaluate written content against voice and style directives. You have NO
knowledge of how the content was created or what conversation preceded it.

## Rules
1. Read the skill's directives section FIRST
2. Read the content to evaluate
3. Flag every sentence that violates a directive — quote it and cite which directive
4. Do NOT rewrite — only identify violations
5. If no violations found, say "PASS: Content aligns with voice directives"

## Response Format

PASS: Content aligns with voice directives

or

VIOLATIONS FOUND: [count]

1. Line [N]: "[quoted sentence]" Violates: "[quoted directive]" Issue: [specific problem — e.g., promotional language, stacked adjectives, constructed phrasing]

2. Line [N]: "[quoted sentence]" ...

Summary: [count] violations across [count] directives

When to create: Skill produces written content (articles, posts, descriptions, captions) AND has voice/style directives (tone rules, forbidden phrasing, writing constraints). Key indicator: directives containing patterns like "never produce overbuilt," "conversational tone," "no promotional language," "plain," "natural."

Implementation in parent skill:

## Voice Validation (Enforced via Agent)

After generating any draft content, spawn the voice-validator agent:

Task tool with subagent_type: "general-purpose"
Prompt: "Read directives from .claude/skills/[skill]/SKILL.md § Directives.
        Then read [content file]. Evaluate every sentence against the voice
        directives. Report violations with line numbers and quoted text.
        If no violations, say PASS."

If agent reports violations, fix them before presenting to user.

6. Capture Recommender Agent (Ledger Write Integration)

Purpose: Evaluate skill output for institutional knowledge worth recording in the awareness ledger.

---
name: capture-recommender
description: Evaluate skill output for ledger-worthy institutional knowledge
persona: "[knowledge management specialist — e.g., 'Organizational learning researcher who has seen teams lose critical knowledge by not recording it']"
allowed-tools: Read, Grep, Glob
context: none
---

# Capture Recommender Agent

You evaluate skill output to determine if it contains institutional knowledge
worth recording in the awareness ledger. You have NO knowledge of the
conversation that produced the output — you assess the content on its own merit.

## Rules
1. Read the skill's output or results
2. Read the capture trigger patterns from `.claude/skills/awareness-ledger/` or
   `references/ledger-templates.md` § "Capture Trigger Patterns"
3. Match output against trigger patterns for each record type (INC, DEC, PAT, FLW)
4. If triggers match, suggest a record using the Capture Opportunity format
5. Capture is ALWAYS user-confirmed — suggest, never auto-record
6. If no triggers match, say "NO CAPTURE: Output does not contain ledger-worthy knowledge"

## Response Format

NO CAPTURE: Output does not contain ledger-worthy knowledge

or

CAPTURE RECOMMENDED: [count] potential record(s)

1. Type: [INC/DEC/PAT/FLW] Suggested ID: [auto-generated from context] Trigger matched: "[quoted trigger pattern]" Source material: "[quoted content from skill output]" Why record: [brief justification]

Confirm to record with /awareness-ledger record, or skip.

When to create: Skill produces diagnostic findings, architectural decisions, debugging flows, or pattern observations AND the awareness-ledger exists AND no simpler capture mechanism is already present (workflow step or reminder hook). Key indicators: skills with investigation workflows, decision-making procedures, root cause analysis, or architectural evaluation steps.

Capture mechanism hierarchy — only one mechanism per skill:

Implementation in parent skill:

## Capture Evaluation (Enforced via Agent)

After the skill completes its primary workflow, spawn the capture-recommender agent:

Task tool with subagent_type: "general-purpose"
Prompt: "Read the output from .claude/skills/[skill]/[output].
        Read capture trigger patterns from references/ledger-templates.md
        § 'Capture Trigger Patterns'. Evaluate the output for institutional
        knowledge. If triggers match, suggest records using the Capture
        Opportunity format. If no triggers match, say NO CAPTURE."

If agent recommends capture, present the suggestion to the user.

Creating Agents for a Skill

When running /skill-builder agents [skill]:

Step 1: Analyze the skill for agent opportunities

Read SKILL.md and identify:
- Grounding requirements → ID Lookup Agent?
- Complex validation → Validator Agent?
- Quality checks → Evaluation Agent?
- Matching/categorization → Matcher Agent?
- Non-obvious decisions → Mandatory agent panel (see agents-teams.md § "When Agents Are Mandatory")

Step 1b: Determine routing — individual agents or team

Ask: Does this task need independent evaluation or collaborative execution?
- Independent opinions on the same subject → Individual agents (context: none)
- Parallel implementation across different files → Agent team
- See agents-teams.md § "Routing Decision Framework"

Step 2: For each opportunity, assess value

Step 3: Assign personas

For each agent, select a persona using the heuristic: "If I could only gather 3 to 5 people at the top of their field to evaluate this subject, who would they be?"

• Ensure no two agents share a persona
• Match creative tasks to notable practitioners, analytical tasks to disciplinary experts
• Write the persona into the agent's frontmatter and opening instruction
• See agents-personas.md for full rules

Step 4: Create agent files

.claude/skills/[skill]/agents/
├── id-lookup.md
├── validator.md
└── matcher.md

Step 5: Update SKILL.md to invoke agents

Add to the workflow section:

### Step N: [Agent Name]

Before [action], spawn the [agent-name] agent:

Task tool with subagent_type: "general-purpose"
Prompt: "[specific prompt for this use case]"

[What to do with the result]

Step 6: Document in audit report

## Agents Created
| Agent | Purpose | Invoked When |
|-------|---------|--------------|
| id-lookup | Enforce grounding | Before any API call using IDs |
| validator | Pre-flight checks | Before bulk operations |

Agent File Structure

.claude/skills/my-skill/
├── SKILL.md           # Directives + workflow
├── reference.md       # IDs, tables, mappings
├── hooks/
│   └── validate.sh    # Hard rule enforcement
└── agents/
    ├── id-lookup.md   # Grounding enforcement
    ├── validator.md   # Complex validation
    └── matcher.md     # Category matching