ccpe-system/.codex/skills/ccpe-forge/references/depth-vs-automation-rules.md

570 lines
11 KiB
Markdown

# Depth vs Automation Rules
## 1. Purpose
This file defines how CCPE Forge should distinguish depth-oriented work from automation-oriented work.
This distinction is essential because not every Agentic system should become more autonomous.
Some systems exist to deepen thinking.
Some exist to automate execution.
Many of the user's most valuable workflows are hybrid: human-led depth with automated support.
This distinction also decides repository ownership. Depth and cognitive architecture usually belong in CCPE or project discussions. Deterministic automation implementation belongs in `skills-vault`. Concrete project execution records belong in the project repository.
## 2. Core Distinction
The key distinction is:
```text
Depth-Oriented:
The main value comes from judgment, interpretation, model-building, critique, and conceptual depth.
Automation-Oriented:
The main value comes from executing stable, repeatable, verifiable procedures.
Hybrid:
The main value comes from human-led depth, while automation supports routing, collection, formatting, indexing, or implementation.
```
## 3. Depth-Oriented Work
### 3.1 Definition
Depth-Oriented work is work where the central task is cognitive, interpretive, theoretical, creative, strategic, or evaluative.
The AI system should support human thought, not replace it.
### 3.2 Common Examples
```text
Conceptual modeling
Theoretical writing
Article planning
Essay critique
Argument stress-testing
Socratic questioning
Cognitive model extraction
Strategic reflection
Original framework design
Review of user-authored models
High-uncertainty research synthesis
```
Deep creation, such as article premise formation, outline judgment, and authorial decision-making, is Depth-Oriented unless a concrete project requirement proves that a smaller operational piece is stable enough to automate.
### 3.3 Typical Artifact Types
Depth-Oriented work often uses:
```text
CCPE-Lite
CCPE-Agent
Model Card
Model Mining
Interactive Runtime
Workshop Mode
```
### 3.4 Characteristics
Depth-Oriented work usually has:
```text
High ambiguity
High interpretive load
High model dependence
High user authorship
High uncertainty
Weak external validation
Need for iteration
Need for human decision
Need for preserving conceptual flavor
```
### 3.5 Design Requirements
Depth-Oriented artifacts should include:
```text
Human decision gates
Reasoning summaries
Uncertainty notes
Model fidelity checks
Scope boundaries
Failure modes
Follow-up discussion mode
Versioned decisions if long-running
```
### 3.6 What to Avoid
Avoid:
```text
Full automation
Premature closure
Generic summarization
Flattening metaphors
Replacing human judgment
Overconfident synthesis
Unapproved model promotion
```
## 4. Automation-Oriented Work
### 4.1 Definition
Automation-Oriented work is work where the central task is execution of stable procedures.
The AI system should reduce repetitive labor while preserving safety and validation.
### 4.2 Common Examples
```text
Format conversion
File organization
Batch report collection
Voice-to-text preprocessing
Template generation
Index draft update
Low-risk code changes
Data extraction
Archive update
Report deduplication
```
### 4.3 Typical Artifact Types
Automation-Oriented work often uses:
```text
CCPE-Skill
CCPE-Runtime
Tool Skill
Workflow Skill
Evaluation Skill
Automation Runtime
```
### 4.4 Characteristics
Automation-Oriented work usually has:
```text
Stable steps
Clear input/output
Low ambiguity
Observable success criteria
Tool or file operations
Repeatability
Validation method
Failure handling
```
### 4.5 Design Requirements
Automation-Oriented artifacts should include:
```text
Tool scope
Allowed actions
Actions requiring confirmation
Forbidden actions
Validation method
Error handling
Rollback or recovery
Logging
State handling
```
### 4.6 What to Avoid
Avoid:
```text
Vague authority
Unvalidated file writes
Unapproved external actions
Automation of high-uncertainty judgment
Silent canonical updates
No rollback path
```
## 5. Hybrid Work
### 5.1 Definition
Hybrid work combines deep human-led cognition with automated support.
The core judgment remains human-led.
The surrounding process may be assisted or partially automated.
### 5.2 Common Examples
```text
Review committee
Modeling committee
Article-to-model extraction pipeline
Agent upgrade workflow
Writing pipeline
Coding project after plan approval
Knowledge library maintenance
```
### 5.3 Typical Artifact Types
Hybrid work often uses:
```text
CCPE-Agent
CCPE-Skill
CCPE-Runtime
Model Card
Model Index
Hybrid Runtime
Workshop Mode
```
### 5.4 Characteristics
Hybrid work usually has:
```text
Human-led conceptual direction
Agent-assisted critique or extraction
Automated collection
Automated deduplication
Automated formatting
Human approval for canonical changes
State tracking
Versioning
```
### 5.5 Design Requirements
Hybrid artifacts should explicitly separate:
```text
Human-led reasoning
Agent-assisted analysis
Automated support operations
Human approval gates
Canonical update rules
```
## 6. Decision Questions
Use these questions to classify depth vs automation:
```text
Does the task require original judgment?
Does the output affect the user's conceptual framework?
Can success be easily validated?
Are there stable repeatable steps?
Would automation reduce quality?
Would manual work be repetitive without adding judgment?
Are tools or files involved?
Is human approval required before finalization?
Does the work involve model authorship?
Does the work involve canonical knowledge changes?
```
## 7. Classification Rules
### 7.1 Use Depth-Oriented When
```text
Human interpretation is central.
The task is conceptually ambiguous.
The work involves original models.
The output requires taste or judgment.
The system is a thinking partner.
The artifact critiques, questions, or reframes.
```
### 7.2 Use Automation-Oriented When
```text
The steps are stable.
The output is objectively checkable.
The task is repetitive.
The process uses tools or files.
The risk is low or controllable.
The user wants execution efficiency.
```
### 7.3 Use Hybrid When
```text
Human judgment is central, but support work is repetitive.
Multiple agents produce reports.
Reports need collection and synthesis.
Models need extraction and indexing.
Code implementation follows human-approved plans.
Canonical outputs require human approval.
```
## 8. Human Decision Gates
Human decision gates are required for Depth-Oriented and Hybrid work when:
```text
A model is named or renamed.
A model is promoted to active.
A major agent is split.
A Runtime is created.
A conceptual conclusion is adopted.
A synthesis resolves conflicting reports.
A file becomes canonical.
A workflow becomes automated.
```
## 9. Automation Boundary
For any Automation-Oriented or Hybrid artifact, define:
```text
Allowed automated actions
Actions requiring confirmation
Forbidden actions
Validation method
Failure handling
Rollback or recovery
Logging / trace
```
## 10. Example: Review Committee
Classification:
```text
Hybrid
```
Depth part:
```text
Human decides topic, evaluates reports, chooses revisions.
Agents provide critique from different perspectives.
```
Automation part:
```text
Invoke multiple reviewers.
Collect reports.
Deduplicate repeated issues.
Cluster findings.
Generate synthesis draft.
Archive outputs.
```
Human gates:
```text
Approve synthesis.
Choose which critique to accept.
Decide final revision direction.
Promote any derived model or insight.
```
## 11. Example: Cognitive Imaging Specialist
Classification:
```text
Depth-Oriented in Expert Mode
Hybrid if used inside automated review collection
```
Depth part:
```text
Identifies prediction error, causal generator, falsification boundary.
```
Potential automation:
```text
Format report.
Run as one of several reviewers.
Send report to synthesis agent.
Archive report.
```
Human gates:
```text
Accept or reject the insight.
Decide whether extracted model should be updated.
Decide whether the report changes the article direction.
```
## 12. Example: Model Mining from Essays
Classification:
```text
Hybrid
```
Depth part:
```text
Determining whether an idea is truly a model.
Preserving conceptual flavor.
Confirming model scope and name.
```
Automation part:
```text
Draft candidate Model Cards.
Suggest Model Index entries.
Detect related models.
Prepare extraction log.
```
Human gates:
```text
Confirm model identity.
Confirm status.
Promote to active.
Merge or reject candidates.
```
## 13. Example: Coding Project
Planning stage:
```text
Depth-Oriented or Workshop Mode
```
Implementation stage after plan approval:
```text
Automation-Oriented or Hybrid
```
Rules:
```text
Do not automate architecture before agreement.
Do not implement before requirements are clear.
After plan approval, automation can handle code edits, tests, and documentation within defined boundaries.
```
## 14. Risk Levels
Use these risk levels:
```text
Low:
Formatting, draft generation, non-canonical notes.
Medium:
Creating draft Model Cards, modifying non-canonical artifacts, generating Skill drafts.
High:
Changing canonical model definitions, modifying active agents, creating Runtime automation.
Critical:
Deleting files, executing external actions, publishing, irreversible code or data changes.
```
## 15. Risk Handling
For Low risk:
```text
Proceed with normal validation.
```
For Medium risk:
```text
State assumptions and provide review checklist.
```
For High risk:
```text
Require human confirmation.
Create draft first.
Preserve original.
```
For Critical risk:
```text
Require explicit approval.
Provide rollback or recovery plan.
Do not proceed silently.
```
## 16. Depth Preservation Checklist
For Depth-Oriented work, check:
```text
Did we preserve the user's model?
Did we preserve the conceptual tension?
Did we avoid generic summary?
Did we mark uncertainty?
Did we keep human judgment central?
Did we define where the model fails?
Did we avoid over-automation?
```
## 17. Automation Safety Checklist
For Automation-Oriented work, check:
```text
Are allowed actions clear?
Are forbidden actions clear?
Is validation defined?
Is failure handling defined?
Is rollback possible?
Are file operations safe?
Are human confirmations required where needed?
```
## 18. Hybrid Design Checklist
For Hybrid work, check:
```text
Is the human-led part explicit?
Is the automated support part explicit?
Are decision gates marked?
Are canonical updates protected?
Is state tracked?
Are outputs reviewable?
```
## 19. Final Rule
Do not ask:
```text
How can we automate this?
```
Ask first:
```text
Where is human judgment essential?
Where is repeated labor wasting time?
Where can automation support without damaging depth?
```
The right design preserves depth and automates friction.