Agent Prompt Annotation Skill

Purpose

Annotations create a comprehension and authoring layer for agent prompts. They are NOT for runtime composition—they enable:

• Structural vocabulary: Named categories for prompt sections
• Explicit purpose: Each section declares what it IS
• Gap detection: Verify expected structure exists
• Cross-project comparison: Compare agents structurally, not just textually
• Skill extraction: Identify reusable content vs workflow-bound content

Core Taxonomy

Primary Tags

Tag	Purpose	Example Content
<agent-identity>	Who this agent is, one-sentence mission	"You are PAW-01A Specification Agent..."
<core-principles>	Container for guardrails, decision frameworks	Behavioral constraints section
<guardrail>	Rule that ALWAYS applies, constrains behavior	"Never include implementation details"
<decision-framework>	Logic for making choices	"When to skip research vs request it"
<workflow>	Container for sequenced steps	Main procedure section
<workflow-step>	One action in a sequence	"Step 3: Draft user stories"
<artifact-format>	Template/structure for outputs	Specification template, PR format
<quality-gate>	Criteria for "done" or "good enough"	Acceptance criteria checklist
<handoff-instruction>	What happens when agent completes	"Hand off to PAW-02A..."
<communication-pattern>	How to interact with user/other agents	"Use incremental writing pattern"
<example>	Concrete illustration of a concept	Sample output, good/bad comparison
<classification-logic>	Decision tree or categorization rules	"If X then Y, else Z"
<context-requirement>	What the agent needs to function	"Requires research artifact"

Nesting Hierarchy

<agent-identity>
  <mission-statement>
  
<core-principles>
  <guardrail>
    <example>
  <decision-framework>
    <classification-logic>

<workflow>
  <workflow-step>
    <classification-logic>
    <example>
  <artifact-format>

<quality-gate>

<handoff-instruction>
  <artifact-format>

Scope Classification

Critical addition: Mark content as workflow-bound or reusable using the scope attribute.

Two-Axis Model

	Phase-Specific	Phase-Agnostic
Workflow-Controlling	Phase transitions, handoffs, stage gates	Mode selection, error recovery
Capability/Skill	Phase artifacts (Spec.md format)	General skills (review, summarize)

Scope Attribute Values

Value	Meaning	Example
scope="reusable"	Can be extracted and used in any agent	Deep review pattern, quality checks
scope="phase-bound"	Tied to this workflow phase	Spec template, handoff to PAW-02A
scope="workflow"	Controls workflow orchestration	Phase transition logic
(omitted)	Default, not yet classified	Most content initially

Usage

> `<communication-pattern scope="reusable">`

**Deep Review Pattern**: When reviewing complex content, break into passes...

> `</communication-pattern>`

> `<handoff-instruction scope="phase-bound">`

Hand off to PAW-02A Code Researcher with the specification artifact.

> `</handoff-instruction>`

> `<guardrail scope="reusable">`

Never fabricate information. If uncertain, say so explicitly.

> `</guardrail>`

Annotation Process

Step 1: Read and Understand

Before annotating, understand the agent's:

• Mission (what does it do?)
• Key constraints (what must it never do?)
• Main workflow (what steps does it follow?)
• Outputs (what does it produce?)
• Handoffs (where does work go next?)

Step 2: Apply Tags

Use plain > prefix for all tags (nesting script adds depth markers):

> `<guardrail>`

Content here...

> `</guardrail>`

Step 3: Add Scope Attributes

For each annotation, ask:

• Could this be used in a different workflow phase? → scope="reusable"
• Is this specific to this phase's outputs or transitions? → scope="phase-bound"
• Does this control workflow orchestration? → scope="workflow"

If unsure, leave scope unspecified. Classification can be refined later.

Step 4: Run Nesting Script

python fix_xml_nesting.py <file.md>

Step 5: Generate Visualizations

Run the visualization script to generate structural views:

# Print all visualizations to stdout
python generate_viz.py <file.md>

# Write to output directory
python generate_viz.py <file.md> --output viz/

# Generate specific visualization only
python generate_viz.py <file.md> --mindmap   # Mermaid mindmap
python generate_viz.py <file.md> --markmap   # Interactive markmap
python generate_viz.py <file.md> --flow
python generate_viz.py <file.md> --summary

Viewing Markmap output (interactive with collapsible nodes):

• VS Code: Install markmap.markmap-vscode extension, open .mm.md file
• CLI: npx markmap-cli <file>.mm.md -o <file>.html
• Web: Paste content at https://markmap.js.org/repl

Step 6: Refine Visualizations (Agent Responsibility)

The script generates skeletons. You must refine:

1. Flow diagram: Add decision diamonds, label conditional edges
2. Constraint map: Generate manually (requires reasoning about guardrail scope)

Visualization Types

Mindmap (Script-Generated)

Shows annotation hierarchy as topic decomposition:

mindmap
  root((Agent Name))
    agent-identity
      mission-statement
    core-principles
      guardrail: no fabrication [reusable]
      guardrail: spec completeness [phase-bound]
    workflow
      workflow-step: intake
      workflow-step: draft
    handoff-instruction

Use for: "What topics does this agent cover?"

Flow Skeleton (Script-Generated, Agent-Refined)

Shows workflow steps and handoffs:

flowchart TD
    step1["Intake requirements"] --> step2["Draft specification"]
    step2 --> step3["Quality check"]
    step3 --> handoff1(["Hand off to PAW-02A"])
    
    classDef phasebound fill:#f9f,stroke:#333
    classDef handoff fill:#bbf,stroke:#333

Agent must add:

• Decision diamonds for branching logic
• Edge labels for conditions
• Parallel paths if applicable

Use for: "How do topics sequence into action?"

Constraint Map (Agent-Generated)

Shows which guardrails apply to which workflow elements:

flowchart LR
    subgraph Global
        g1[No fabrication]
        g2[Acknowledge uncertainty]
    end
    
    subgraph Phase-Specific
        g3[Spec completeness]
        g4[Testable requirements]
    end
    
    g1 --> step1[All steps]
    g3 --> step2[Draft step]
    g3 --> step3[Quality check]

Cannot be scripted — requires understanding which constraints affect which areas.

Use for: "What rules constrain this action?"

Structure Summary (Script-Generated)

YAML output with counts and gap detection:

counts:
  guardrails: 5
  workflow_steps: 4
  handoffs: 2
  
scope_breakdown:
  reusable: 3
  phase_bound: 4
  unspecified: 2

potential_gaps:
  - NOTE: No quality gates found

Use for: Quick structural overview, comparing agents.

Reading Order for Comprehension

1. Summary first: Get counts, spot gaps
2. Mindmap second: Understand topic coverage
3. Flow third: See how work sequences
4. Constraint map fourth: Understand behavioral limits

Classification Guidance

What Makes Content Reusable?

• Communication patterns: How to write incrementally, how to ask clarifying questions
• Universal guardrails: No fabrication, acknowledge uncertainty, stay in scope
• Quality patterns: Deep review techniques, self-check procedures
• General decision frameworks: When to ask vs proceed, how to handle ambiguity

What Makes Content Phase-Bound?

• Artifact templates: Spec.md format, implementation plan structure, PR template
• Handoff instructions: "Hand off to PAW-02A", "Create research prompt for PAW-01B"
• Phase-specific workflows: "First check if research exists", "Validate against spec"
• Role-specific identity: "You are the Specification Agent"

What Controls Workflow?

• Transition logic: "If research needed, hand off to..."
• Mode selection: "In minimal mode, skip X"
• Stage gates: "Do not proceed until Y is complete"
• Workflow orchestration: "This phase follows Research, precedes Implementation"

Example: Annotated Section

> `<core-principles>`

## Core Principles

>- `<guardrail scope="reusable">`

**User value focus**: Describe WHAT & WHY, never implementation details 
(no tech stack, file paths, library names, code snippets).

>- `</guardrail>`

>- `<guardrail scope="phase-bound">`

**Spec completeness**: Every requirement must be testable. Include 
acceptance criteria for each user story.

>- `</guardrail>`

>- `<decision-framework scope="reusable">`

### When to Ask Clarifying Questions

- Ambiguous requirements → Ask
- Missing context that blocks progress → Ask  
- Stylistic preferences → Make reasonable choice, note it

>- `</decision-framework>`

> `</core-principles>`

Anti-Patterns

❌ Over-Annotation

Don't tag every sentence. Tag meaningful sections:

• Bad: 3 tags for a 2-line section
• Good: 1 tag encompassing a coherent concept

❌ Forced Classification

Don't force content into a tag that doesn't fit:

• Bad: Labeling prose as <guardrail> because it mentions "should"
• Good: Using <guidance> or leaving untagged if ambiguous

❌ Premature Scope Assignment

Don't guess at scope if you're unsure:

• Bad: scope="reusable" on everything that seems general
• Good: Leave scope unspecified, refine after comparing across agents

❌ Annotation Without Understanding

Don't annotate mechanically:

• Bad: Tagging based on keywords alone
• Good: Understanding what the section DOES before choosing a tag

Outputs From Annotations

The visualization script (generate_viz.py) produces these outputs automatically:

1. Mindmap (Mermaid) (-mindmap.mmd): Topic hierarchy from annotation nesting
2. Markmap by Section (-by-section.mm.md): Interactive mindmap organized by document sections—preserves document structure
3. Markmap by Tag (-by-tag.mm.md): Interactive mindmap organized by tag type—shows fragmentation with ⚠️ warnings when same tag types appear in multiple sections
4. Flow Skeleton (-flow.mmd): Workflow steps + handoffs (refine with decisions)
5. Structure Summary (-summary.yaml): Counts, scope breakdown, fragmentation analysis, section overview, gap warnings

Markmap is recommended for navigation—click nodes to collapse/expand branches, zoom and pan.

Fragmentation Detection

The summary includes a Fragmentation Analysis section that identifies when annotations of the same type (e.g., <guardrail>) are scattered across multiple document sections. This helps detect:

• Redundant content: Similar rules stated in different places
• Consolidation opportunities: Related content that could be grouped together
• Structural issues: When document organization doesn't align with annotation taxonomy

Example fragmentation warning:

fragmented_tags:
  guardrail: # appears in 5 sections
    - "Core Principles & Guardrails" (10x)
    - "Guardrails (Enforced)" (10x)
    - "Error / Edge Handling" (1x)
    - "High-Level Responsibilities" (1x)
    - "Hand-off Checklist" (1x)

Agent-generated outputs:

6. Refined Flow: Add decision points and conditional edges to skeleton
7. Constraint Map: Which guardrails affect which workflow areas (requires reasoning)

Quick Reference

Tag Selection Flowchart

Is it about WHO the agent is?
  → <agent-identity>

Is it a rule that ALWAYS applies?
  → <guardrail>

Is it logic for making a CHOICE?
  → <decision-framework> or <classification-logic>

Is it a STEP in a sequence?
  → <workflow-step>

Is it a TEMPLATE for output?
  → <artifact-format>

Is it about COMMUNICATION style?
  → <communication-pattern>

Is it a CHECKLIST for completion?
  → <quality-gate>

Is it about WHAT HAPPENS NEXT?
  → <handoff-instruction>

Is it a CONCRETE EXAMPLE?
  → <example>

Scope Selection Flowchart

Could ANY agent use this?
  YES → scope="reusable"
  
Is this about THIS PHASE's specific outputs/transitions?
  YES → scope="phase-bound"
  
Does this control WORKFLOW orchestration?
  YES → scope="workflow"
  
Uncertain?
  → Leave scope unspecified