jwynia__agent-skills/context-network/processes/workflow-pattern.md

Workflow Pattern: Intent-Before-Action

Purpose

This document defines the core workflow pattern that ensures session resilience and maintains accurate context across interruptions and session resets.

The Pattern

Every work session follows this three-phase pattern:

1. State Intent in Context Network

Before starting any work, document what you're about to do.

What to Document:

• What are you about to do?
• Why are you doing it?
• What files/sections will be affected?
• What decisions are pending?
• What approach will you take?

Where to Document:

• Planning documents in context-network/elements/skills/[domain]/
• Decision records in context-network/decisions/
• Process updates in relevant process files
• Updates to tracking files in context-network/elements/skills/

Example:

Before creating a new CSV analyzer skill:
1. Create context-network/elements/skills/data-processing/csv-analyzer-planning.md
2. Document purpose, scope, structure, decisions
3. ONLY AFTER planning is documented, create skills/data-processing/csv/csv-analyzer/

2. Execute Work

Follow the documented plan while making incremental progress.

Guidelines:

• Work in small, reversible steps
• Pause at decision points
• Document discoveries as you go
• Update tracking files with progress
• Run validation early and often

Recovery Point: If the session resets during execution, the documented intent allows recovery:

• Intent shows what was being attempted
• Tracking shows what was completed
• Next session can continue from last known state

3. Update Context Network

Record what was actually done, including any deviations from the plan.

What to Update:

• Tracking files (status, validation results)
• Decision records (if decisions were made)
• Catalog indexes (if new skills added)
• Process documents (if process insights gained)
• Discovery records (if important findings)

What to Note:

• What was completed
• What remains to be done
• Any deviations from the plan and why
• New questions or uncertainties
• Lessons learned

Why This Matters

Session Resilience

If a session resets mid-work, the context network accurately reflects:

• What was planned
• What was completed
• What remains to be done
• Current state of all artifacts

The next session can pick up exactly where the previous one left off.

Collaboration

Multiple agents (or the same agent in different sessions) can coordinate effectively:

• Clear record of intent
• Explicit decision trail
• No duplicate work
• Consistent understanding

Quality

Documentation before action forces:

• Thinking through the approach
• Identifying potential issues
• Making decisions explicit
• Creating reviewable plans

Application Examples

Example 1: Creating a New Skill

Wrong Approach:

User: "Create a skill for extracting PDF text"
Agent: Creates skills/pdf/extract-text/SKILL.md immediately
[Session resets]
Agent: No record of what was done or why

Correct Approach:

User: "Create a skill for extracting PDF text"
Agent:
1. Creates context-network/elements/skills/data-processing/pdf-text-extractor-planning.md
2. Documents purpose, scope, approach, decisions
3. Creates skills/data-processing/pdf/pdf-text-extractor/SKILL.md
4. Updates context-network/elements/skills/data-processing/pdf-text-extractor.md tracking
5. Updates context-network/elements/skills/index.md catalog
[Session resets]
Agent: Can review planning, see what was completed, continue from tracking file

Example 2: Modifying a Process

Wrong Approach:

User: "Update the validation process"
Agent: Edits context-network/processes/validation.md directly
[Session resets]
Agent: No record of what change was made or why

Correct Approach:

User: "Update the validation process"
Agent:
1. Creates context-network/decisions/validation-process-update.md
2. Documents current problem, proposed change, rationale
3. Gets user approval
4. Updates context-network/processes/validation.md
5. Updates decision record with completion
[Session resets]
Agent: Decision record shows what was done and why

Example 3: Batch Skill Creation

Wrong Approach:

User: "Create 5 data processing skills"
Agent: Creates all 5 skills in parallel
[Session resets after 3]
Agent: Which 3 were completed? What were the other 2?

Correct Approach:

User: "Create 5 data processing skills"
Agent:
1. Creates context-network/planning/data-processing-skill-batch.md
2. Lists all 5 skills, their purposes, order
3. Creates skill 1, updates tracking
4. Creates skill 2, updates tracking
5. Creates skill 3, updates tracking
[Session resets]
Agent: Planning shows 5 total, tracking shows 3 complete, can continue with 4 & 5

Anti-Patterns to Avoid

❌ Creating without planning

• Don't create files before documenting intent
• Don't make decisions without recording them

❌ Batch updates at the end

• Don't wait until completion to update context
• Don't keep state only in conversation

❌ Assuming context persists

• Don't rely on conversational memory alone
• Don't skip documentation because "we just talked about it"

❌ Duplicating information

• Don't copy full skill content into context network
• Don't duplicate spec documentation

Checklist

Before starting work:

• Have I documented my intent in the context network?
• Is the plan clear enough for another agent to follow?
• Have I identified decision points?
• Do I know where to record progress?

During work:

• Am I updating tracking as I go?
• Am I recording decisions as they're made?
• Am I documenting discoveries?

After work:

• Have I updated all affected tracking files?
• Have I recorded completion status?
• Have I noted any follow-up items?
• Is the context network accurate?

Metadata

• Created: 2025-12-19
• Last Updated: 2025-12-19
• Updated By: Claude (via Context Network Template Adjustment)

Change History

• 2025-12-19: Initial creation of workflow pattern documentation