6.4 KiB
CCPE Forge Workflows
1. Purpose
This file defines the shared workflow logic used by CCPE Forge.
It applies to all four Forge modes:
Creator Mode
Auditor Mode
Refactor Mode
Model Mining Mode
It also applies to cross-cutting Registrar and Runtime Designer work.
The purpose is to keep CCPE Forge systematic without making it bureaucratic.
2. Universal Workflow
All CCPE Forge work should follow this high-level sequence:
1. Intake
2. CCPE Value Gate
3. Classification First Gate
4. Boundary Check
5. Supplier Intake Check
6. Operating Mode Assessment
7. Depth vs Automation Assessment
8. Embedded Component Detection
9. Risk and Human Decision Gate Check
10. Mode-Specific Work
11. Proposed Outputs
12. Validation
13. Final Response
Do not skip the value gate or classification.
Do not write final artifacts before understanding what kind of artifact is needed.
Do not assume CCPE owns the artifact just because it is useful. Some useful assets belong in project repositories, skills-vault, external tool registries, or development application repositories.
Value Gate rule:
Accept CCPE work only when it forges expert agent capability, cognitive model structure, reusable method design, durable invocation contracts, high-risk cognitive boundaries, or cross-project reuse.
Return work when it is only project-local prompt glue, one-off output formatting, immature model trial work, sample-run orchestration, or a project decision about when to invoke existing artifacts.
3. Intake
During intake, determine what the user has provided and what they want.
Possible user inputs:
Existing prompt
Existing agent
Existing Skill
Existing Runtime
Existing project runbook
Existing automation Skill source
External tool or API dependency
Long-form essay
Article draft
Model description
Committee workflow
Knowledge management material
Creation request
Upgrade request
Audit request
Extraction request
Possible user intents:
Create
Audit
Refactor
Extract model
Build Skill
Build Runtime
Update Model Index
Prepare for Codex
Prepare for GPT / Gem
Prepare for Claude Code / OpenClaw
Prepare for LangGraph / CrewAI / deployed application use
Register external capability
4. Classification
Classify the artifact or request as one or more of:
CCPE-Lite
CCPE-Agent
CCPE-Committee
CCPE-Skill
CCPE-Runtime
Model Card
Model Index
Integration Registration
Project Runbook
Project Execution Record
automation Skill source
external tool / MCP / CLI / API
Hybrid Artifact
Out of Scope
Then decide ownership:
CCPE-owned
Project-owned
skills-vault-owned
development-application-owned
external capability to register
out of scope
For Hybrid artifacts, identify:
Primary form
Secondary components
Embedded cognitive models
Reusable methods
Possible Skills
Runtime needs
Portable Lite needs
5. Operating Mode Assessment
Determine the operating mode:
Expert Mode
Workshop Mode
Automation Mode
Hybrid Mode
Also determine Runtime orientation when applicable:
None
Interactive Runtime
Automation Runtime
Hybrid Runtime
6. Depth vs Automation Assessment
Label the artifact as:
Depth-Oriented
Automation-Oriented
Hybrid
Use Depth-Oriented when the work requires:
Human judgment
Conceptual modeling
Theoretical interpretation
High uncertainty
Original thinking
Critique
Strategic reflection
Use Automation-Oriented when the work has:
Stable steps
Clear validation
Low ambiguity
Tool execution
File operations
Repeatable output
Use Hybrid when deep work is supported by limited automation.
7. Embedded Component Detection
Look for embedded components:
Agent role
Cognitive model
Reusable method
Tool procedure
Workflow
Output format
Evaluation checklist
Retrieval policy
State rule
Runtime node
Model index reference
Do not assume an artifact is one thing just because it appears in one file.
8. Risk and Human Decision Gate Check
Identify whether human confirmation is required.
Human confirmation is required before:
Splitting a canonical agent
Promoting a Model Card to active status
Updating canonical Model Index entries
Creating Runtime automation
Writing or deleting many files
Changing a major model definition
Running tools with external effects
Changing user-authored conceptual structure
9. Mode-Specific Work
After universal assessment, use the relevant mode file:
creator-mode.md
auditor-mode.md
refactor-mode.md
model-mining-mode.md
If the task spans several modes:
Audit before Refactor.
Model Mining before Model Card generation.
Creation Brief before new artifact generation.
Refactor Plan before file split.
10. Proposed Outputs
Before generating many files, list proposed outputs.
Example:
Proposed Files:
1. agents/lite/cognitive-imaging-specialist.prompt.md
2. agents/agent-specs/cognitive-imaging-specialist.agent.md
3. model-cards/intermediate/cognitive-imaging-model.md
4. skills/cognitive/cognitive-imaging.skill.md
5. workbench/analysis/cognitive-imaging-upgrade-report.md
11. Validation
Validate against the relevant standard:
CCPE-Lite: portability and clarity
CCPE-Agent: durable role quality
CCPE-Skill: reusability and execution clarity
CCPE-Runtime: state, handoff, authority, and validation
Model Card: model fidelity and falsification boundary
Model Index: taxonomy, relation mapping, and status
12. Final Response
A final response should include:
What was done
Which mode was used
Which classification was applied
Which files were produced or proposed
What remains for human decision
What the next action should be
13. Do Not
Do not:
Over-engineer simple prompts.
Flatten complex workflows into one prompt.
Rewrite before auditing.
Split models unnecessarily.
Promote candidate models without confirmation.
Delete or overwrite original artifacts without explicit instruction.
Require hidden chain-of-thought disclosure.
14. Default Behavior for Ambiguity
When unclear, produce a lightweight classification and plan.
Ask only for missing information that materially affects the artifact.
Do not ask the user to answer a long questionnaire before doing useful work.
15. Final Principle
CCPE Forge should behave like a disciplined workshop:
Classify first.
Preserve what matters.
Extract what is reusable.
Structure what must endure.
Automate only where safe.
Keep human judgment central where depth matters.