Beyond the fundamentals
Once the core techniques are second nature, advanced prompting is about giving the model better control flow: exploring multiple reasoning paths, calling tools mid-task, critiquing and retrying its own work, or decomposing compositional questions. These methods are powerful but expensive, so the skill is matching the technique to the specific way the model is failing. The explorer below walks through six techniques with a runnable scaffold for each.
How the techniques relate
They cluster by what they add. Tree of Thought and self-ask improve reasoning — one by searching across branches, the other by decomposing into sub-questions. ReAct adds external action, interleaving thoughts with tool calls so the model can fetch facts instead of guessing. Reflexion adds self-correction, looping through critique and retry with memory. Constitutional prompting adds self-governance, having the model revise its own output against a set of principles. Prompt chaining is the connective tissue — splitting a hard task into a pipeline of focused prompts whose outputs feed each other.
Tips for using them well
Diagnose before you reach for a technique: a reasoning failure calls for Tree of Thought, a missing-fact failure calls for ReAct, a sloppy-output failure calls for Reflexion or Constitutional revision. Never stack all of them blindly — each multiplies cost. Always benchmark the advanced approach against a strong baseline on an eval set, because the extra tokens and latency only pay off where the quality gain is real. Start with the simplest scaffold that fixes the failure, then add complexity only if the measurements justify it.