| name | meta-methodology-extraction |
| description | Use when extracting the implicit methodology from a deep-dive conversation — reverse-engineering the analytical path and operational principles into a reusable Framework + Workflow + SOP. Triggers include "extract the methodology", "derive a framework from this conversation", "what method did we use here", "turn this into a repeatable process", "提煉出方法論". Do NOT use for summarizing content; this extracts the method, not the subject. |
| version | 0.2.0 |
Meta-Methodology Extraction
Extract the implicit methodology from a deep-dive conversation. Output three files under .agents-stack/methodology-extraction/<id>/: framework.md, workflow.md, sop.md, validation-report.md.
This is not a fill-in-the-blank template. The methodology must be discovered from the conversation, not projected onto it. The steps below describe what to observe, not what to produce.
What to Observe
When reading the source conversation, look for:
1. Layer Progression
How did the analysis move from start to finish? Identify the natural stages — do NOT force a predefined layer model. Common patterns include:
- Broad taxonomy → pick one problem → deep-dive mechanism → root cause → solutions → vision
- Phenomenon → symptom → cause → fix → production norms
- Current state → why broken → what's being done → what should be done → how to do it
Key question: If you had to teach someone to analyze a similar problem, what stages would you tell them to go through, in what order?
2. Principles
What operational rules governed the analysis? These are not domain facts ("LLMs are probabilistic") but analytical rules ("if a symptom can't be measured, it can't be diagnosed"). Look for:
- Heuristics the analyst applied (e.g., "always ask for evidence before accepting a claim")
- Boundary rules (e.g., "stop drilling when you hit a structural invariant")
- Quality criteria (e.g., "a symptom must be observable, not interpretive")
3. Cognitive Moves
What structural devices did the analysis use repeatedly?
- Comparisons (A vs B tables)
- Level separation (training layer / inference layer / engineering layer)
- Time horizons (now / 3-5 years / ultimate)
- Model mapping (applying known frameworks to new domains)
- Priority ordering (what to build first and why)
Key question: If you removed these devices, would the analysis still hold its structure?
Extraction Process
Step 1: Read for Path
Read the conversation end-to-end. Do not take notes. Just absorb the arc.
Then answer: "This conversation moved from ____ to ____ to ____ to ____." This is your analytical path — the backbone of the methodology.
Step 2: Extract the Framework
From the path, derive:
- Mental models: The core ways of seeing that made the analysis possible. 3-5, each with a plain-language explanation and source in the conversation.
- Taxonomies: If the conversation created classification systems, extract them. Do NOT invent taxonomies that weren't in the conversation.
- Principles: The operational rules behind each stage of the path. Every principle must answer "why this stage matters."
- Structural patterns: The devices used to organize knowledge. Catalog them.
For every element: is it universal (would work on a different domain) or domain-specific (tied to the conversation's topic)?
Step 3: Build the Workflow
Translate the path into executable steps. Each step needs:
- What: the action
- Why: the principle behind it (from Framework)
- Input: what must be available before starting
- Output: what must be produced before moving on
The workflow must be executable by someone who never read the source conversation.
Step 4: Derive SOP
From the Workflow, extract operational norms:
- Rules: when to start, when to stop, when to redo a step
- Quality gates: objective completion criteria per step
- Metrics: how to tell if the process is healthy
- Anomaly responses: what to do when a step fails or stalls
Step 5: Validate
Back-test everything against the source conversation:
- Coverage: does the Framework capture every significant pattern?
- Gaps: what was in the conversation but NOT in the extraction? Why?
- Over-extraction: are any one-time observations being mistaken for universal rules?
- Boundary: which parts are universal, which are domain-specific, which are provisional?
Output Structure
.agents-stack/methodology-extraction/<id>/
├── framework.md # Path, mental models, taxonomies, principles, structural patterns
├── workflow.md # Executable steps: what, why, input, output
├── sop.md # Rules, quality gates, metrics, anomaly responses
└── validation-report.md # Coverage, gaps, over-extraction flags, boundary statement
Output Format (Consolidated Response)
After extraction is complete, report:
- Depth verdict: sufficient / insufficient
- Analytical path: "This conversation moved from ____ to ____ to ____ to ____."
- Framework summary: key mental models, principle count
- Workflow summary: step count
- Validation: coverage gaps, over-extraction flags, boundary classification
Critical Rules
- Never force a predefined structure onto the conversation. The methodology must emerge.
- Never confuse domain content (what was discussed) with methodology (how it was discussed).
- Every extracted element must be traceable to the source conversation.
- The workflow must be standalone — executable without reading the source.
- If the conversation is too shallow to extract methodology, say so and stop. Do not fabricate.
