Task and output contract
Turn a request into a job with a clear deliverable and a clear way to check it.
81 methods · Three levels in every card
Prompt Atlas
A starting point based on the work you want to do.
Search for a method name or for the work you want to do. When you open a card, you can read the explanation, the Simple, Medium and Hard examples, the limits and the original sources together.
Start with the Simple example; read its situation and limits before adapting the prompt. The examples are fictional teaching examples, not model experiments that were actually run.
Read the guide to choosing a method →
Card texts: Mustafa Saraç · CC BY 4.0. Review / contribute on GitHub. Source version: 6257da10f016. Images and third-party works are not covered by this license.
How this image was made
Original illustration made with Google Gemini · 1024 × 572
Generate an image: Create an original museum-grade computational sculpture photographed as a physically present installation, horizontal 16:9. Deep anthracite void, mineral porcelain whites, restrained ice-blue and amber accents, fine particles only where structurally meaningful, tactile surfaces, subtle volumetric illumination, believable depth, exceptional edge detail and intentional negative space. A monumental asymmetric mineral atlas suspended in darkness: four unmistakably different physical pathways share one quiet white origin. One pathway extracts fragments from deep layered strata, one folds a broad sheet into an expressive surface, one passes through exact geometric joints, and one forms a responsive open loop reaching into the environment. The pathways remain spatially distinct and meet at a small controlled amber nexus. View from a low three-quarter angle with architectural scale and dramatic but soft side lighting. Preserve a large dark breathing area in the upper left. The result should feel like a single intentional sculpture, not four panels or a collage. Keep the image sophisticated and legible at mobile size: one dominant mechanism, a clear silhouette, no decorative overload. No text, letters, numbers, typography, captions, symbols, watermark, logo, fake chart, screenshot, interface, identifiable people or artist signature. Do not imitate any specific existing artwork.
Explanations, examples and sources are searched too. If you type several words, all of them must match.
Showing 81 of 81 cards. All use cases.
Few-shot
Show the behavior you want with a few correct input–output pairs.
Role and persona
State which details the model should look at, and for which reader.
CO-STAR
Make the context, objective and reader visible in the same short brief.
Chain-of-Thought
Carry the short, checkable intermediate results from a solved example over to a new problem.
Zero-shot CoT
Ask for short intermediate results without giving a solved example; extract the final answer separately.
Least-to-Most
Solve the easy subproblem first; make its result the input of the next problem.
Step-Back
Before returning to the detailed question, find which general principle applies.
Self-Consistency
Take separate samples for the same question; count the final answers in a common format.
Tree of Thoughts
Open more than one solution path; evaluate the intermediate states and go back when needed.
PAL
Let the model write the program; let a real interpreter compute the numeric answer.
RAG — Find the source, ground the answer in it
When you answer a question, find the relevant document first; then build the answer from the information in that document.
Citation contract
Match every claim with the passage that actually supports it.
HyDE
Generate a draft that resembles the document you are looking for; use it to find the real documents.
IRCoT
Use the intermediate information you found to change the direction of the next search.
Structured Outputs
Give the output schema to an API that supports it; check the meaning of the returned data separately.
Delimiters and labeled context
Separate the instruction, the source text and the example in a readable way.
Self-Refine
Pass the first text through a separate critique call and correct it with concrete feedback.
Chain-of-Verification
Break the draft into small check questions; get the answers again without being influenced by the draft.
Reflexion
Turn the feedback from one attempt into a short lesson note and carry it into the next attempt.
LLM-as-a-judge
Get the model's evaluation with explicit criteria and sources; check the judge's decision too.
Meta prompt that writes prompts
Ask the model not to do the job, but to write the instruction that will be used for that job.
Meta-Prompting: expert calls
Let a conductor model define the subtasks; let the application pass them to separate expert calls.
Meta Prompting: structural schema
Instead of solved content examples, give a schema that shows the structure of the task.
Automatic Prompt Engineer
Generate candidate instructions; choose which one works based on the target model's real outputs.
OPRO
Generate new candidates by showing earlier candidates together with their real scores.
GEPA
Derive prompt changes from error traces; keep candidates that are strong on different examples together.
DSPy
Define model calls as input–output modules; tie optimization to a metric.
Inverse Prompting
Rank candidates by measuring, as a probability, how well the generated text predicts the starting prompt.
ReAct
Choose the next step based on the tool result that actually came back.
Prompt chaining
Split a job into small calls whose output is checked.
Tool and result contract
Define what comes back from a tool as carefully as what goes into it.
Bounded execution and permissions
Make the permissions, the budget and the stop condition visible at the start of the job.
Modular master prompt
Keep a project's reusable working rules in one document.
Context compaction
When you shorten a long history, do not lose the decisions and the open tasks.
Structured external notes
Keep the state that needs to be remembered outside the chat, in a note with sources.
Flipped Interaction
Before preparing the answer, let the model ask for the missing information.
Question Refinement
Before answering a question, propose a more useful version of it.
Uncertainty and abstention
Make clear where the answer ends and which information is missing.
Separating untrusted input
Do not confuse commands inside the text being read with the user's authority.
RUNE
Turn a scattered request into a layered instruction structure with a known version.
Chain of Draft
Instead of narrating intermediate steps at length, limit them to short, checkable notes.
Graph of Thoughts
Work with solution pieces that can merge, instead of a single path.
Generated Knowledge
Before answering, generate candidate pieces of relevant knowledge; do not mistake them for sources.
Plan-and-Solve
First set up the solution plan, then carry out the steps of that same plan.
Self-Ask
Answer the missing subquestion before returning to the main question.
EmotionPrompt
Emotional emphasis is an instruction variable; it does not give the model emotions.
Skeleton-of-Thought
First produce a skeleton, then fill in the independent parts in separate calls.
System 2 Attention
First rewrite the context relevant to the question; generate the answer from that context.
Chain-of-Density
Bring in the important information left in the source without making the summary longer.
Directional Stimulus Prompting
Let a small policy model generate task-specific hints for a large model.
Medprompt
Select similar examples, shuffle the options, and combine separate answers by mapping them back.
CRITIC
Tie the critique to an external tool result, then correct the answer.
Thread of Thought
Examine a scattered context in small parts, then extract only the answer.
Program of Thoughts
Hand the numerical calculation over from the language model to an executor.
Reversing Chain-of-Thought
Reconstruct backwards which question the solution assumes, and compare it with the actual question.
Active Prompting
Have a human label the examples the model is most uncertain about.
Decomposed Prompting
Route subtasks to named handlers; carry their results like a program.
Rubber Duck
Look for the missing point by explaining what you want to do and which steps you follow.
Self-explanation
You build the explanation; the model makes the connection you skipped visible.
Teach-back
Explain back what you understood in your own words; check where the explanation fell short.
Teachable agent
Let the model keep the student role with a limited knowledge state; you are the one explaining.
Socratic tutoring dialogue
Without giving the answer right away, help the learner find the next step.
Mutual argument questioning
Question each other's assumptions in turn; keep the final decision open.
Recursive Socratic Questioning
Split a subquestion again when needed; combine the answers upwards.
Retrieval practice
Recall without seeing the answer, then compare with the source.
Spaced repetition
Spread review over time; record what you forgot.
Problem framing
Before choosing a solution, clarify which problem you are solving.
Eliciting preferences
Do not ask for a preference as a single label; bring it out through trade-offs.
Self-Debug
Have the generated code explained, check how it runs, fix the faulty step.
Metacognitive Prompting
Make understanding the task, the first solution and the final check separate, visible steps.
Cognitive Verifier
Before building the main answer, clarify the necessary subquestions with the user.
Alternative approaches
Put other paths next to the first solution; compare them with the same criteria.
Analogical Prompting
Generate similar solutions, then show which structure carries over to the target question.
Joint ideation
Run ideation together with the human's own ideas and explicit constraints.
Premortem
Assume the work has failed; look for concrete causes that could lead to it.
Argument map
See the claim, its support and its objection as separate nodes.
Human judgment first
Record your own judgment first, then compare it with the model's suggestion.
Facilitating group discussion
Make the group's different pieces of knowledge visible; do not manufacture an artificial consensus.
Multiagent Debate
Compare the answers of separate model calls, then run a bounded debate.
Tutor supervisor
The tutor produces the answer; a separate supervisor checks the permitted level of help.
Adapt the prompts to your own situation. In an example that needs a tool or a separate call, copying the text alone does not set up that way of working.