Advanced prompting
Advanced prompting is the set of techniques you reach for when a clear, structured prompt is not enough: self-consistency, few-shot design, constraint calibration, tree of thoughts, meta-prompting, evaluation rubrics and prompt compression. They matter most on multi-step reasoning, long documents and work someone will act on. It does not mean jailbreaks.
Start with the basics — a role, the situation, one task, rules and an output format (the RCTCO structure). The techniques below build on that. Each one is taught in a short course and used inside the advanced items of the library.
The techniques, what they fix, where to learn them
| Technique | What it does | Learn it |
|---|---|---|
| Chain-of-thought, done right | Makes the model reason in checkpoints before it answers, instead of narrating its first guess. | Advanced CoT Mastery · glossary |
| Self-consistency | Samples several independent answers and keeps the one most of them agree on — for problems with one right answer. | Self-Consistency Voting · glossary |
| Few-shot design | Teaches a pattern with two or three examples instead of a paragraph of description; picking the examples is the skill. | Few-Shot Pattern Recognition · glossary |
| Constraint calibration | Finds the few rules that change the output and drops the ones the model ignores or over-applies. | Constraint Calibration |
| Tree of thoughts | Explores several lines of reasoning, scores them and continues the best — for planning and open problems. | Tree-of-Thoughts for Planning · glossary |
| Meta-prompting | Has the model write or improve the prompt itself, against criteria you set. | Meta-Prompting · glossary |
| Evaluation rubrics | Scores outputs against written criteria, so a change to a prompt is judged on evidence rather than a feeling. | Prompt Evaluation Rubrics |
| Agents (ReAct) | Lets the model reason, call a tool, look at the result and continue — with a stop condition. | Building Simple Agents · glossary |
| Prompt compression | Cuts a long brief to the words that carry the instruction — cheaper, and often clearer on small models. | Compression skill |
In what order to learn them
- Structure first: role, context, one task, rules, output format. Most “advanced” problems are missing context.
- Then examples (few-shot) and a reasoning step (chain-of-thought) — the two with the widest effect.
- Then checking: a rubric, a self-critique pass, or several samples compared (self-consistency).
- Then planning and agents (tree of thoughts, ReAct) — for work that takes more than one step.
- Last, compression: shorten what works, and keep the shorter version only if it scores as well.
Advanced items in the library
- Tree-Of-Thoughts Strategy Explorer For AgentsHas an agent branch multiple solution strategies, evaluate each, prune weak branches, and commit to the strongest path.Plan
- Few-Shot Grounded QA PatternUses worked examples to teach the model the exact grounded-answer and refusal behavior expected.Plan
- Agent Evaluation Rubric And Trace GraderCreates an objective rubric and grades an agent execution trace on task success, tool use, efficiency, and safety.Plan
- Few-Shot Example DesignerCreate optimal training examples that teach LLMs patterns through demonstration rather than instruction.Plan
- Prompt Compression & Token OptimizerCuts prompt token cost while preserving behavior, using lossless restructuring and instruction prioritization.Plan
13 of the 36 courses are free — start with Basic Chain-of-Thought. The advanced courses are on Pro; the advanced library items are on the Library and Pro plans. See the plans.
Questions
What is advanced prompting?
The techniques for tasks a single clear prompt does not solve: self-consistency, few-shot design, constraint calibration, tree of thoughts, meta-prompting, evaluation rubrics and compression. They build on a structured prompt; they do not replace it.
Is advanced prompting the same as jailbreaking?
No. Jailbreaking tries to get a model to break its rules. Advanced prompting gets better, more checkable work out of a model within them — clearer reasoning, better examples, stricter output, evidence that a change helped.
Which advanced technique should I learn first?
Few-shot examples and a reasoning step. They help on the widest range of tasks. Add a checking step — a rubric or a self-critique pass — as soon as someone will act on the output.
Do these techniques work on every model?
The ideas carry across ChatGPT, Claude and Gemini, but the size of the effect varies by model and task. That is why the evaluation-rubrics course matters: measure a change on your own examples before you keep it.