146 lines
3.2 KiB
Markdown
146 lines
3.2 KiB
Markdown
# Refactor Plan: Cognitive Imaging Specialist
|
|
|
|
## 1. Scope
|
|
|
|
```text
|
|
source_path: workbench/raw/认知显影者1.1.md
|
|
current_phase: Auditor Mode only
|
|
refactor_status: proposed, not executed
|
|
```
|
|
|
|
This plan does not upgrade or split the artifact yet. It only defines the recommended direction for a later Refactor Mode pass.
|
|
|
|
## 2. Recommended Target Form
|
|
|
|
```text
|
|
Recommended Target: Hybrid split, preserving a Lite version
|
|
```
|
|
|
|
Suggested future structure:
|
|
|
|
```text
|
|
1. Model Card
|
|
-> 认知显影术 / Cognitive Imaging
|
|
|
|
2. CCPE-Skill
|
|
-> cognitive-imaging five-step method
|
|
|
|
3. CCPE-Agent Spec
|
|
-> Cognitive Imaging Specialist role and authority
|
|
|
|
4. CCPE-Lite Prompt
|
|
-> portable one-piece version for chat products
|
|
|
|
5. Optional Runtime node
|
|
-> only if used in review committee or model-mining workflow
|
|
```
|
|
|
|
## 3. Refactor Sequence
|
|
|
|
### Step 1. Preserve original
|
|
|
|
Keep `workbench/raw/认知显影者1.1.md` unchanged.
|
|
|
|
### Step 2. Extract Model Card draft
|
|
|
|
Extract only the cognitive model itself:
|
|
|
|
```text
|
|
model name
|
|
aliases
|
|
source material
|
|
scope
|
|
core assumptions
|
|
mechanism
|
|
procedure
|
|
inputs
|
|
outputs
|
|
failure modes
|
|
falsification boundary
|
|
related skills
|
|
related agents
|
|
runtime usage
|
|
version status
|
|
```
|
|
|
|
Recommended status: `draft`, not `active`.
|
|
|
|
### Step 3. Extract Skill draft
|
|
|
|
Convert the five-step method into a reusable CCPE-Skill:
|
|
|
|
```text
|
|
Trigger:
|
|
complex adaptive system analysis
|
|
|
|
Procedure:
|
|
Capture
|
|
Darkroom
|
|
Enlarger
|
|
Exposure
|
|
Development
|
|
|
|
Validation:
|
|
prediction error identified
|
|
at least one causal test performed
|
|
prohibition line stated
|
|
external data treated as raw material
|
|
```
|
|
|
|
### Step 4. Convert role into Agent Spec
|
|
|
|
The Agent should contain:
|
|
|
|
```text
|
|
objective
|
|
role
|
|
context
|
|
capability
|
|
authority
|
|
workflow
|
|
constraints
|
|
state
|
|
output
|
|
evaluation
|
|
collaboration
|
|
```
|
|
|
|
The Agent should reference the Model Card and Skill rather than embedding all of them.
|
|
|
|
### Step 5. Preserve portable Lite version
|
|
|
|
Keep a compact version for GPT / Gemini / Claude custom assistant usage. This is important because the original artifact is already effective as a standalone expert prompt.
|
|
|
|
### Step 6. Propose Model Index update
|
|
|
|
Do not promote to active automatically. Keep or update status as:
|
|
|
|
```text
|
|
status: draft
|
|
review_status: needs-user-confirmation
|
|
```
|
|
|
|
## 4. Human Decision Gates
|
|
|
|
Before Refactor Mode writes upgraded artifacts, the user should confirm:
|
|
|
|
```text
|
|
1. Is the canonical Chinese model name "认知显影术"?
|
|
2. Should the role name be "认知显影专家" or "认知显影者"?
|
|
3. Should Cognitive Imaging be a draft Model Card first, or active after review?
|
|
4. Should the portable Lite version preserve the original severe voice?
|
|
5. Should联网检索 be allowed as tool use, or only as externally supplied context?
|
|
6. Should this Agent become a committee member later?
|
|
```
|
|
|
|
## 5. Refactor Risks
|
|
|
|
- Over-splitting could weaken the original prompt's immediacy.
|
|
- Over-sanitizing could destroy the darkroom / film-development metaphor.
|
|
- Promoting the Model Card too early could freeze an incomplete model.
|
|
- Treating retrieval as autonomous tool use without policy could exceed user intent.
|
|
|
|
## 6. Recommended Next Step
|
|
|
|
Run Refactor Mode only after the user confirms the target file set and the canonical naming decisions.
|