Skip to content
Multigrid

Prompt engineering

Which prompting techniques actually change a model's output, which ones better models absorbed, and how to tell the difference on your own workload.

Prompt engineering acquired a bad reputation honestly. A great deal of what was published between 2022 and 2024 was folklore: phrases that worked once, on one model, for reasons nobody isolated, repeated until they became advice. Then the models improved and most of it stopped doing anything.

What is left is narrower and more durable. A prompt is the entire program — there is no state between calls — so everything the model can use has to be in it, arranged so the model can find it, in a shape you can check in code. These pages are about that: what to put in, where to put it, what to stop putting in, and how to know which is which without trusting anyone’s list of tricks, including this one.

Prompt Engineering: What Still Works in 2026

The techniques that survived three generations of better models, the ones the models absorbed, and the order to try them in.

5 min read

System Prompts: What Belongs There and What Does Not

A placement rule for every kind of instruction, derived from what actually differs between the system role and the user role.

5 min read

Few-Shot Prompting: How Many Examples Is Enough?

What examples actually teach a model, what the published literature says about the shape of the accuracy curve, and the exact cost of every extra shot.

5 min read

Chain of Thought Is Now Built In. Should You Still Prompt for It?

When an explicit reasoning instruction still earns its output tokens, and when it duplicates work the model is already billing you for.

5 min read

Role Prompting: Does "You Are an Expert" Do Anything?

What the published persona studies found, why the mechanism predicts exactly that result, and the part of role prompting that does still work.

4 min read

Delimiters, XML Tags and Structuring a Long Prompt

How to separate instructions from data so a model can tell them apart, and why each family documents a different convention.

5 min read

Negative Instructions: Why "Don't" Often Backfires

Why prohibitions are a weak instruction form, ten real ones rewritten into positives, and what to enforce in code instead.

5 min read

Prompt Chaining vs One Big Prompt

A cost model for decomposition you can run on your own token counts, and the error arithmetic that decides the accuracy side.

5 min read

Self-Consistency: Sampling Five Times and Voting

The majority-vote technique, the gains its authors reported, and the break-even calculation that tells you whether five times the cost is worth it for you.

5 min read

ReAct: Interleaving Reasoning and Tool Calls

The loop behind almost every agent, as the original paper defined it and as it looks now that tool calling is an API feature.

5 min read

Meta-Prompting: Using an LLM to Write Your Prompts

The optimiser loop, the published results that make it credible, and the statistics that explain why most of its wins evaporate.

5 min read

Prompt Templates and Variables Without a Framework

Why string interpolation breaks on prompts specifically, and a hundred-line templating layer that fails loudly instead.

6 min read

Versioning Prompts Like Code

A git-based workflow where every prompt change is a diff, every generation is attributable to a version, and the eval runs in CI.

5 min read

Why the Same Prompt Behaves Differently on Every Model

The nine ways a working prompt breaks when you point it at another model, each with the symptom you will actually see first.

4 min read

Prompt Compression Without Losing Accuracy

What to cut from a prompt first, ordered by how recoverable the loss is, and the harness that tells you when you have cut too far.

6 min read

Instruction Following: Why Models Ignore Rule 7 of 12

Why compliance degrades as rules accumulate, how to measure it on your own rule set, and what to do instead of adding rule 13.

5 min read

Output Formatting: Consistent Shapes Without JSON Mode

Getting parseable output from models with no structured-output support: format choice, prompt-side tactics, tolerant parsing and a retry that works.

5 min read

Prompting for Refusals: Making a Model Say "I Don't Know"

Abstention as a designed output rather than a hoped-for behaviour: give it a literal, make it evidence-backed, and measure both directions.

5 min read

A/B Testing Prompts in Production

Splitting traffic between two prompts, sizing the experiment before you run it, and the analysis mistakes that make prompt tests unusually easy to get wrong.

5 min read

The Prompt Engineering Techniques That Stopped Mattering

A graveyard with causes of death: what the models absorbed, what the APIs replaced, and what was never real to begin with.

4 min read

Other topics