Reasoning Graph Generation

Reasoning Graph Generation

Reference for: Common Ground Load when: Using --graph flag, generating mermaid diagrams

Purpose

The reasoning graph makes Claude's decision-making structure visible—not just the assumptions (premises), but the decision tree that led to the current approach.

Two complementary artifacts:

• COMMON-GROUND.md = the premises (what we're assuming)
• Reasoning graph = the structure (how decisions connect)

Mermaid Diagram Structure

Basic Template

flowchart TD
    %% Root node: the task/goal
    ROOT[Task: {task_description}]

    %% Decision points: diamond shape
    ROOT --> D1{Decision Question?}

    %% Branches with weights and source tags
    D1 -->|"weight: 0.8 [inferred]"| P1[Chosen Path]
    D1 -->|"weight: 0.2 [alternative]"| P2[Alternative Path]

    %% Downstream decisions
    P1 --> D2{Next Decision?}

    %% Leaf nodes: concrete implementations
    D2 --> I1[Implementation Detail]

    %% Styling
    style D1 fill:#ffcc00,stroke:#333
    style P1 fill:#90EE90,stroke:#333
    style P2 fill:#cccccc,stroke:#333

Node Types

Styling Rules

Apply styles after all node definitions:

%% Decision points: yellow
style D1 fill:#ffcc00,stroke:#333
style D2 fill:#ffcc00,stroke:#333

%% Chosen paths: green
style P1 fill:#90EE90,stroke:#333

%% Alternatives: gray
style P2 fill:#cccccc,stroke:#333

%% Uncertain: orange
style U1 fill:#FFB366,stroke:#333

%% Implementations: blue
style I1 fill:#87CEEB,stroke:#333
style I2 fill:#87CEEB,stroke:#333

Edge Labels

Include weight and source tag on edges where relevant:

D1 -->|"weight: 0.8 [stated]"| P1
D1 -->|"weight: 0.5 [inferred]"| P2
D1 -->|"[uncertain]"| U1
D1 -->|"[alternative]"| A1

Source Tags

Node Naming Conventions

IDs

Use consistent prefixes:

Labels

• Keep labels concise (< 30 chars)
• Use sentence case
• Decision points end with ?
• Implementations are specific

Graph Generation Process

Step 1: Identify Root Task

Extract the main task/goal from conversation context:

ROOT[Task: Build authentication system]

Step 2: Map Major Decisions

For each ESTABLISHED or WORKING assumption, trace back:

• What decision led to this assumption?
• What alternatives existed?

Step 3: Build Decision Tree

Connect decisions hierarchically:

ROOT --> D1{Scope?}
D1 --> P1[Production] --> D2{Architecture?}
D1 --> A1[MVP]
D2 --> P2[Stateless] --> D3{Token Strategy?}
D2 --> A2[Stateful]

Step 4: Add Leaf Implementations

For concrete decisions (ESTABLISHED assumptions):

D3 --> I1[JWT access tokens]
D3 --> I2[Refresh token rotation]
I1 --> I3[15min expiry]
I2 --> I4[7-day refresh window]

Step 5: Mark Uncertainty

Flag OPEN assumptions and uncertain branches:

D2 -->|"[uncertain]"| U1[Redis sessions?]
style U1 fill:#FFB366,stroke:#333

Step 6: Apply Styling

Add style rules for all nodes based on their type.

Complete Example

Context

User wants to build an authentication system. Through conversation:

• Production-grade confirmed (inferred from requirements)
• Stateless preferred (assumed, not confirmed)
• JWT chosen over sessions (working assumption)

Generated Graph

flowchart TD
    ROOT[Task: Build auth system] --> D1{MVP or Production?}
    D1 -->|"0.8 [inferred]"| P1[Production-grade]
    D1 -->|"0.2 [alternative]"| A1[MVP/Prototype]

    P1 --> D2{Stateless required?}
    D2 -->|"0.7 [assumed]"| P2[JWT + refresh tokens]
    D2 -->|"0.3 [uncertain]"| U1[Redis sessions]

    P2 --> I1[Access token: 15min]
    P2 --> I2[Refresh token: 7 days]
    P2 --> I3[Token rotation on refresh]

    U1 --> I4[Session in Redis]
    U1 --> I5[Cookie-based ID]
    U1 --> D3{Scaling concern?}
    D3 -->|"[uncertain]"| U2[Redis cluster needed]
    D3 -->|"[alternative]"| A2[Single instance OK]

    style D1 fill:#ffcc00,stroke:#333
    style D2 fill:#ffcc00,stroke:#333
    style D3 fill:#ffcc00,stroke:#333
    style P1 fill:#90EE90,stroke:#333
    style P2 fill:#90EE90,stroke:#333
    style A1 fill:#cccccc,stroke:#333
    style U1 fill:#FFB366,stroke:#333
    style U2 fill:#FFB366,stroke:#333
    style A2 fill:#cccccc,stroke:#333
    style I1 fill:#87CEEB,stroke:#333
    style I2 fill:#87CEEB,stroke:#333
    style I3 fill:#87CEEB,stroke:#333
    style I4 fill:#87CEEB,stroke:#333
    style I5 fill:#87CEEB,stroke:#333

Embedding in COMMON-GROUND.md

Add the graph as a section in the ground file:

---

## Reasoning Graph

Last generated: {timestamp}

```mermaid
flowchart TD
    ...

Graph Legend

---

## File Output Options

### Option 1: Embedded (Default)

Add `## Reasoning Graph` section to `COMMON-GROUND.md`

**Pros:** Single file, travels with assumptions
**Cons:** Larger file

### Option 2: Standalone

Create `REASONING.mermaid` in project root

**Pros:** Can be viewed in mermaid-compatible editors
**Cons:** Two files to track

### Option 3: Both

Embed in `COMMON-GROUND.md` AND create standalone file

**When:** User explicitly requests standalone file

---

## Conversational Updates

### Expanding Branches

User: "Expand the MVP branch"

1. Find the `A1[MVP/Prototype]` node
2. Add downstream decisions for that path
3. Regenerate with expanded subtree

### Switching Paths

User: "I actually want Redis sessions"

1. Swap styling: `U1` becomes green, `P2` becomes orange
2. Update weights on edges
3. Expand `U1` subtree if needed
4. Regenerate graph

### Adding Detail

User: "What's downstream of token rotation?"

1. Find `I3[Token rotation on refresh]` node
2. Add child nodes for implementation details
3. Regenerate with expanded leaf

---

## Validation Rules

### Graph Must Have

- Exactly one ROOT node
- At least one decision point (D node)
- Styling for all nodes
- Edge labels with weights or tags

### Graph Must Not Have

- Orphan nodes (disconnected)
- Cycles (this is a tree, not a graph)
- Missing style definitions
- Unlabeled decision edges

---

## Regeneration Triggers

Regenerate the graph when:

| Trigger | Action |
|---------|--------|
| Assumption promoted | Change node color to green |
| Assumption demoted | Change node color to orange |
| New assumption added | Add branch or leaf node |
| Alternative explored | Expand grayed branch |
| User requests expansion | Add subtree detail |

Always preserve the full tree structure—never remove alternatives, just gray them out.