Ingénierie de graphe
Laying the work out around the model as an explicit graph instead of one long chat. Three rungs. **Prompt engineering** improves the question. **Context engineering** improves the information. **Graph engineering** improves the circuit: one request splits into parallel branches, each with its own role and exit criterion, then merges back into a single recommendation you sign off before acting. What separates it from a long conversation is not model power, it is **traceability**: every branch has a readable mandate, and you can see which one produced what.
Strengths
- Every branch has a readable role and exit criterion, making it traceable which one produced what
- Independent branches move in parallel, where a conversation stays sequential by nature
- The merge point forces a human sign-off before action, instead of continuous execution
Limitations
- Cost multiplied by the number of branches, for zero gain if the work is linear
- A bad split produces redundant branches the merge will have to arbitrate with no criterion
- The term pulls you toward tooling when the useful work is drawing, not installing
Best for
- A product call that must cross several independent angles before deciding
- A review where the drafter and the verifier must remain separate models
- Long work where you want to say afterwards which step produced which conclusion