Most prompt engineering advice in 2026 is cargo cult. It gets copied from a viral thread, restated in a listicle, and passed around until it feels like consensus, without anyone checking whether the underlying study was three models stale or built on fifty hand-written questions. "Be rude to the model." "Add urgency." "Minimize context or it rots." These read like insight. They fall apart on contact with evidence.

So we did the boring thing and checked. We ran an adversarial deep-research pass: 109 agents, multi-source fetching from primary vendor docs and peer-reviewed papers, and a three-vote refutation on every single claim. A technique had to survive verification to make the "works" list. Several popular ones did not survive at all. Here is the honest split.

The techniques that actually work

These cleared the bar. Notice the pattern: they are all about prompt architecture and specification, not about clever wording or coaxing the model into a mood.

Put your long data at the top, your question at the bottom

For long-context tasks (roughly 20k tokens and up), Anthropic's current docs are explicit: place long documents and inputs near the top of your prompt, above your query and instructions. Their docs note query placement "can improve response quality by up to 30 percent in tests, especially with complex, multidocument inputs."

We are labeling that 30 percent honestly, because you should too: it is an unaudited first-party vendor figure with no published methodology. A verifier on our team traced it back and found it does not cleanly match Anthropic's older long-context post (which reported a 36 percent error reduction from a different scratchpad strategy on much older models). So do not repeat "30 percent" as a proven fact. What is solid is the direction, and it is independently corroborated by the "lost in the middle" position-bias literature (Liu et al.; arXiv 2406.16008), which documents the same U-shaped attention curve and confirms it persists in 128K-plus context models. Data at the top, primary command at the very bottom. That part is real.

Use 3 to 5 diverse examples to steer format, not to teach reasoning

Few-shot prompting is not dead, but its evidenced job is narrower than people think. Anthropic's docs call examples "one of the most reliable ways to steer Claude's output format, tone, and structure" and suggest 3 to 5 diverse examples. Crucially, examples steer shape, not intelligence. "Calibrate Before Use" (arXiv 2102.09690) found accuracy can actually degrade as you pile on examples, with up to 30 points of variance from example ordering alone, and random labels barely hurting. Examples act as format specifiers, not task teachers.

One honest scope note the listicles skip: this dosage is provider-specific. OpenAI's guidance for its reasoning models says try zero-shot first, then add examples only if needed. So "always use 3 to 5 examples" is wrong on OpenAI's o-series and GPT-5. Know your model.

Structure with XML tags, but never attach a percentage to it

Wrapping different content types in their own tags (<instructions>, <context>, <input>) genuinely helps the model parse mixed prompts. Anthropic endorses it. But the mechanism is disambiguation, not a measured accuracy uplift. We ran a targeted falsification grep across the full best-practices page for any benchmark number attached to XML tags and found zero. The one "30 percent" on the page belongs to data placement, not tags. Use XML structuring. Do not let anyone sell it to you with a fake stat.

Nail structured extraction with a real recipe

The strongest commercial use case is structured extraction, and OpenAI's GPT-5.2 guide gives a concrete, testable recipe: always supply a schema, use strict structured outputs, distinguish required from optional fields, set missing fields to null rather than guessing, and add a re-scan pass to catch omissions. One caveat that ships with it: grammar-constrained decoding guarantees valid shape, never correctness, and can degrade reasoning by roughly 10 to 15 percent (JSONSchemaBench, arXiv 2501.10868). The fix is to free-reason first, then format.

Let the LLM plan and a tool execute

For anything computational, the ICLR 2025 meta-analysis (Sprague et al., arXiv 2409.12183) is decisive: an LLM that plans and hands execution to a tool (Python, a solver) dominates prompt-based reasoning. The LLM still does the planning. You are not bypassing it, you are giving it a calculator.

The myths that failed the vote

These are the ones to stop repeating. Each failed our three-vote refutation threshold.

Myth: politeness, rudeness, or tips change the answer

Refuted. The well-powered study here is Cai et al. (arXiv 2512.12812, December 2025): 1,446 questions, three tones, ten repetitions, three models, nearly 130,000 responses. The finding: "the majority of 95 percent confidence intervals include zero." Only 5 of 18 task-level comparisons reached significance, all in humanities, and the largest effect found anywhere was 3.22 percentage points. Gemini 2.0 Flash showed no significant effect on any comparison.

The viral "be rude and get 84.8 percent" claim you have seen? That was one model on fifty hand-written questions. The better-powered study finds less effect, which is exactly what you expect if the original number was noise. Tone is not where prompt value lives.

Myth: manual chain-of-thought is a default best practice

Refuted as a default; survives only as a scoped fallback. Anthropic itself demotes manual chain-of-thought to a fallback for when thinking is off, and states a general instruction like "think thoroughly" often beats a hand-written step plan. The peer-reviewed evidence is blunt: the ICLR 2025 meta-analysis found chain-of-thought delivers +14.2 percent on symbolic tasks and +12.3 percent on math, but only +0.7 percent across everything else (56.8 percent versus 56.1 percent direct). On the exact business tasks most people prompt for (extraction, classification, brand-voice writing, summarization), it is removable cost, not a quality lever. It still helps for math, logic, and cases where you need auditable, separated reasoning. It is not a universal on-switch.

Myth: context rot means you should always minimize context

Refuted, 0 to 3. The idea that models have a fixed attention budget so trimming context is inherently a performance technique is a widely repeated practitioner meme that did not survive a single vote. Position and structure matter. "Less context is always better" is not supported.

Myth: response prefilling to force JSON, and emphatic pressure prompting

Both refuted, 0 to 3. "Anthropic recommends prefilling the response to enforce JSON/XML" failed 0 to 3. So did "emphatic pressure prompting (CRITICAL: You MUST...) is now counterproductive." Note what that second refutation means honestly: it does not prove threats work, it means there is no surviving first-party evidence either way. Stop citing it as settled.

The uncomfortable meta-point

Here is the part nobody selling a "secret prompt pack" wants to say out loud: the techniques that survived are all public. They are in Anthropic's docs, OpenAI's cookbook, and open-access papers on arXiv. There is no secret phrase. The taxonomy is even catalogued (Schulhoff et al.'s "The Prompt Report" lists 58 text-based techniques), but a taxonomy existing is not the same as those techniques being validated for your task.

Which is the whole point. Because the methods are public, the real value is not knowing them. It is measuring them on your own data. The 30 percent figure is unaudited. The extraction recipe is vendor-recommended but not independently benchmarked on your documents. Every one of these techniques has a scope limit, a provider caveat, or an effect that only shows up on certain task types. The only way to know what works for your workload is a before-and-after eval you actually run.

Why we ship prompt systems with an eval harness

This is exactly why our prompt-skills catalog does not sell magic phrases. Each system ships with the verified technique applied to a specific task (extraction, classification, long-document analysis, brand-voice writing) and an eval harness so you can measure the before-and-after on your own inputs rather than trust a vendor percentage you can trace to a dead end. It is $25. You own it, you run the eval, you keep the numbers. No cargo cult, no borrowed stats.

The honest summary of 2026 prompt engineering is this: the wording tricks got refuted, the architecture and specification techniques held up, and the only figure worth trusting is the one you measured yourself.