Skip to content

methodologies.md

Latest commit

 

History

History
192 lines (137 loc) · 7.2 KB

methodologies.md

File metadata and controls

192 lines (137 loc) · 7.2 KB

Methodologies Guide

Methodologies are reasoning frameworks that the server injects into your prompts. They structure how the LLM thinks — adding phase labels, quality criteria, and evaluation patterns to every response.

Why This Matters

Problem Solution Result
Unstructured output Phase-based reasoning (Context, Analysis, Goals...) Consistent, reviewable sections
Skipped thinking Methodology gates with validation criteria Each phase checked for depth
One-size-fits-all 6 built-in + custom creation Match the framework to the task

Tip

Quick start: Run system_control(action: "framework", operation: "switch", framework: "cageerf") to activate a methodology. Every prompt after that receives CAGEERF phase guidance automatically.

Built-in Methodologies

ID Name Best For Phases
cageerf C.A.G.E.E.R.F General-purpose structured reasoning Context, Analysis, Goals, Execution, Evaluation, Refinement
react ReACT Iterative reasoning + action loops Reason, Act, Observe (repeat)
5w1h 5W1H Investigative analysis Who, What, When, Where, Why, How
scamper SCAMPER Creative ideation and brainstorming Substitute, Combine, Adapt, Modify, Put to other use, Eliminate, Reverse
focus FOCUS Problem-solving with root cause analysis Find, Organize, Clarify, Understand, Solution
liquescent LIQUESCENT Creative flow and artistic exploration Layered creative phases

Using Methodologies

Switch the Active Framework

The active methodology applies to all subsequent prompts in the session:

system_control(action: "framework", operation: "switch", framework: "cageerf")

Per-Request Override

Use the @ operator to apply a specific methodology to a single prompt without changing the session default:

prompt_engine(command: "@REACT >>my_prompt")

Disable for a Single Request

Use modifiers to suppress methodology injection:

Modifier Effect
%clean Disable all injection (methodology + gates + style)
%lean Disable methodology and style, keep gates
%guided Force all injection on
%framework Force methodology injection on 
prompt_engine(command: "%clean >>my_prompt")

Note

Standalone prompts (non-chain) inject the methodology once. For chains, injection frequency is configurable — see Injection Control.

What Gets Injected

When a methodology is active, the server adds guidance at multiple levels:

Layer What Where
System prompt guidance Phase descriptions and reasoning instructions Prepended to system prompt
Methodology gates Per-phase quality validation criteria Added to gate review
Tool description overlays Framework-branded tool descriptions Visible in MCP tool listing
Phase guards Structural assertions on output sections Post-execution verification
Example: CAGEERF system prompt injection

When CAGEERF is active, this guidance is injected into every prompt:

Apply the C.A.G.E.E.R.F methodology systematically:

**Context**: Establish comprehensive situational awareness and environmental factors
**Analysis**: Apply structured, systematic examination of the problem or opportunity
**Goals**: Define specific, measurable, actionable objectives with clear success criteria
**Execution**: Develop practical, implementable approach with detailed action steps
**Evaluation**: Create robust success metrics and assessment methods
**Refinement**: Enable continuous improvement and iteration processes

Creating a Custom Methodology

Use the built-in >>create_methodology prompt to design and validate a new framework:

prompt_engine(command: ">>create_methodology", options: {
  "name": "My Framework",
  "concept": "A methodology for systematic API design"
})

The prompt guides you through designing phases, then validates against a 5-tier completeness score (100% required):

Tier Weight What It Checks
Foundation 30% id, name, system prompt guidance, phases (min 2)
Quality 20% Methodology gates with validation criteria
Authoring 25% Required sections, argument suggestions, template hints
Execution 15% Processing steps with assertions, execution steps with dependencies
Advanced 10% Tool description overlays, quality indicators, judge prompt

Tip

Study the CAGEERF definition at server/resources/methodologies/cageerf/ for a complete reference implementation covering all 5 tiers.

File Structure

Each methodology lives in its own directory under server/resources/methodologies/:

server/resources/methodologies/{id}/
├── methodology.yaml    # Configuration, gates, guidance, tool overlays
├── phases.yaml         # Phase definitions, processing steps, assertions
└── judge-prompt.md     # Optional: judge evaluation prompt for %judge modifier

Files are hot-reloaded — edit and save, the server picks up changes automatically.

Phase Guards (Assertions)

Processing steps can include section_header + guards for deterministic output verification. After execution, the server checks the LLM's response against these rules — no LLM cost, instant feedback:

processingSteps:
  - id: context_establishment
    name: Context Establishment
    order: 1
    required: true
    section_header: "## Context"
    guards:
      required: true
      min_length: 100
      forbidden_terms: ["TODO", "TBD", "placeholder"]
Rule Type Description
required boolean Section must exist in the response
min_length number Minimum character count
max_length number Maximum character count
contains_any string[] Must include at least one term
contains_all string[] Must include all terms
matches_pattern string Regex the section must match
forbidden_terms string[] Terms that must NOT appear

Note

Phase guards are separate from gates. Guards check structure (did the LLM produce a ## Context section?). Gates check quality (is the context analysis thorough?). See Phase Guards Guide for details.

Checking Status

system_control(action: "status")

Returns the active framework, available methodologies, and current injection settings.

system_control(action: "framework", operation: "list")

Lists all registered methodologies with their enabled status.

See Also

  • Injection Control — Frequency, targets, and modifiers for methodology injection
  • Phase Guards — Deterministic structural validation of methodology phases
  • Gates Guide — Quality validation criteria (complements methodology structure)
  • Judge Mode — Context-isolated evaluation using methodology judge prompts
  • Architecture Overview — How methodologies integrate with the pipeline