| name | deltas |
| description | Extract deltas from the project vision |
| disable-model-invocation | true |
Delta Extraction Workflow
Extract deltas from VISION.md into DELTAS.md.
Context
You must load the following skills and read the following files before proceeding.
Skills
katachi:framework-core - Workflow principles
Required files
docs/planning/VISION.md - Project vision to extract deltas from
docs/planning/DELTAS.md - Current delta inventory (if exists)
General Guidance
Follow the collaborative workflow principles from the framework-core skill.
Deltas-specific guidance:
Use a scratchpad - Track state in /tmp/deltas-<project-name>-state.md (where <project-name> is the basename of the current working directory):
- Raw deltas identified with source traceability
- Agent validation findings
- Refinements to apply
Extract all at once - Review entire vision and extract all deltas in one pass.
Detect gaps proactively - Challenge completeness:
- "What handles errors here?"
- "What validates this input?"
- "Should we consider security/logging/configuration?"
Process
0. Check Existing State
If docs/planning/DELTAS.md exists:
- Read current delta inventory
- Read current vision (check for changes)
- Compare: Are there new workflows? Changed scope?
- Ask: "Should we review for new deltas, or refine existing ones?"
- Enter iteration mode as appropriate
If no deltas exist: proceed with initial extraction
1. Extract All Deltas at Once
Review the vision document thoroughly. Extract all deltas in a single pass:
Consider all aspects:
- User-facing capabilities (what users can DO)
- Core workflows and interactions
- Foundational user needs
- Cross-cutting user concerns (security, privacy, error handling from user perspective)
Each delta represents a bounded piece of work:
- Can be a new capability, modification, or improvement
- Should be implementable in a reasonable timeframe
- May affect one or more areas of the system
- Description must follow the format in the DELTAS template (
references/DELTAS-template.md from framework-core)
For each delta, document source traceability (which vision section).
2. Agent Validation #1: Raw Delta List
Dispatch the delta-validator agent to review the raw delta list.
Task(
subagent_type="katachi:delta-validator",
prompt=f"""
Validate this raw delta list.
## Vision Document
{vision_content}
## Raw Delta List
{raw_deltas}
"""
)
Review agent findings with user.
3. User Iteration on Raw Deltas
Based on agent validation:
- Split deltas that are too large or deliver multiple capabilities
- Merge or remove redundant deltas
- Add missing deltas identified
- Ensure each delta is traceable to vision
4. Organize and Number
Assign sequential IDs (DLT-001, DLT-002, etc.) to all deltas.
For each delta:
- Show: proposed ID, description, complexity
- User reviews each delta
- Validate/adjust as needed
5. Agent Validation #2: Complete Delta Inventory
Dispatch the delta-validator agent to review the completed inventory.
Task(
subagent_type="katachi:delta-validator",
prompt=f"""
Validate this complete delta inventory.
## Delta Inventory
{deltas_md_content}
"""
)
Review agent findings with user.
6. User Iteration and Finalization
Ask: "Should we iterate based on this feedback, or is the inventory complete?"
If gaps to address → refine deltas
If complete → finalize and write to docs/planning/DELTAS.md
7. Dependency Analysis
After the delta inventory is finalized, analyze all deltas for dependencies.
Analyze each delta
For each delta, analyze:
- What data/capabilities does it consume from other deltas?
- What other deltas must exist for this to work?
- What does it share with other deltas (state, configuration)?
- What must be initialized before this can start?
- What must work for this to be testable?
Apply dependency priority principles:
- Core capabilities before workflows
- Core workflows before enhancements
- Foundation before configuration
- Functionality before polish
- Configuration deltas are enhancements, not prerequisites
Present proposed dependencies
Present all proposed **Depends on**: values with reasoning for each delta.
User iteration on dependencies
User reviews proposed dependencies. Discuss and adjust based on user knowledge.
Surface hidden dependencies with gap detection questions:
- "Does X need data that Y produces?"
- "What must be initialized before X can start?"
- "Do X and Y access the same configuration?"
- "What must work for X to be testable?"
- "If Y fails, does X need to know?"
Agent validation
Dispatch a general-purpose subagent to validate the dependency graph.
Include user's explicit decisions made during analysis.
Request critique on:
- Missing implicit dependencies
- Hidden coupling (shared config, state, resources)
- Over-specification (unnecessary dependencies)
- Circular dependencies (cycles)
If cycles found, work with user to resolve:
- Re-examine if dependency is really needed
- Revise delta scope
- Split deltas
Review findings with user. Iterate if needed.
Write dependencies inline
Write dependencies as **Depends on**: DLT-XXX, DLT-YYY (or None) in each delta's entry in docs/planning/DELTAS.md.
Workflow
This is a collaborative process:
- Extract deltas systematically
- Challenge gaps and completeness
- User confirms complexity, IDs
- Challenge deltas that are too large or too small
- Never add deltas without user agreement
- Analyze and propose dependencies after inventory is finalized
- Iterate until complete