Organize complex graphs

Large graphs stay useful when branches have clear purposes and visible boundaries. Convography provides generated titles and branch colors; your prompts and branching choices provide the structure.

Branch at meaningful decisions#

Create branches where assumptions, audiences, scenarios, or methods diverge. Avoid branching every minor follow-up. A graph with five meaningful scenario roots is easier to read than one with twenty nearly identical retries.

Name the difference in the first prompt on a branch. Generated node titles can then reflect the distinction.

Use branch colors deliberately#

Open a node menu and choose a color to override the automatic branch accent. The chosen color follows that branch visually. Use Auto color to return it to the default palette.

One simple scheme is:

• green for the main or recommended path;
• pink for risk and counterargument;
• purple for research;
• yellow for open questions;
• blue for implementation.

Colors do not change context or priority. Document your scheme in an early context node if other people will use the graph.

Combine mature branches#

After exploring alternatives, use Synthesize to select the text or context endpoints you want to combine. Convography includes each selected endpoint's complete ancestor path, includes shared history once, and creates one assistant-only result with an incoming edge from every selected endpoint. Ordinary follow-up messages still use only their single active ancestor path; sibling branches are included only when you explicitly select them for synthesis.

Split mature work#

When one branch develops its own audience or long sequence, use Split conversation. This keeps the source graph readable while preserving the complete path in a new conversation.

Use files as visible evidence#

Files used by graph nodes can appear as connected file nodes. Give uploaded files descriptive names before using them, select only relevant context, and inspect each response’s Included Files list. The graph should make evidence relationships clearer, not turn into an undifferentiated file dump.

Current scale limits#

The current interface supports pan and zoom but does not yet include subtree collapse, graph search, persistent viewport state, or virtualization designed for thousands of nodes. For now, split independent work before a single graph becomes unwieldy.