Over the past 18 months, we've built 545 hand-crafted prompt Originals on Promptolis — across 21 categories, in two languages (EN + DE), each shipped with a complete example output of 2,000-4,000 words. We've also indexed thousands of community-contributed prompts from awesome-chatgpt-prompts and curated public sources.
This is the most comprehensive labeled dataset of professional-grade prompts in existence (we know — we looked).
Here's what the data shows about how prompt engineering actually works in 2026.
This isn't a "how to write a better prompt" article. This is what 545 hand-crafted prompts told us about which structural patterns produce useful outputs, which produce smooth nothing, and which categories of work AI is genuinely changing — versus which categories are just AI-decorated.
Section 1: What 545 prompts look like in aggregate
Before we get into the patterns: the dataset.
- Total Originals: 545
- Categories covered: 21 (from Career & Work to Spiritual & Lifestyle)
- Average prompt length: 1,847 characters (XML structure + principles + input + output-format)
- Average example output length: 2,790 words
- Tier distribution: Tier 1 (volume keywords) 350 · Tier 2 (sweet spot) 106 · Tier 3 (gold/moat) 89
- Most-populated category: Wellness & Health (62 Originals)
- Least-populated category: Spiritual & Lifestyle (15 Originals)
The category distribution isn't random — it reflects two forces: (1) where the keyword-volume opportunity is largest, and (2) where AI prompts genuinely produce useful work versus where they decorate.
We expected Career & Work to be the largest category. It's not — it's the 14th. Wellness & Health is. The reason: wellness questions are highly bounded and repeatable ("audit my burnout signals," "diagnose my habit failure"), and the cost of generic advice is real but bounded. Career questions are higher-stakes but more idiosyncratic — fewer questions can be solved by frameworks alone.
Section 2: The 7 Structural Patterns That Recur in Working Prompts
Across 545 prompts, the patterns that produce reliably-good outputs are surprisingly consistent. We've documented these before in our original 7-patterns analysis from 336 prompts — here's the updated 545-prompt version.
Pattern 1: Specific role + experience anchor
```
You are a [role] with [N] years of [specific] experience. You have [specific quantified background — "edited 80+ mysteries" or "worked with 1,000+ families"].
```
Why it works: Models trained on the internet recognize patterns from professional voice. "Expert" produces general advice. "12-year negotiation strategist who's worked both sides of executive comp" produces calibrated advice.
Failure mode if absent: Generic "act as expert" prompts produce the median voice — useful 60% of the time, useless 40%, and consistently lacking the specificity that makes a prompt feel professional.
Frequency in our 545: 100%. Every working Original opens with this pattern.
Pattern 2: Principles section (numbered, contrarian, specific)
```xml
- [Specific principle — not platitude]
- [Contrarian insight from the field]
- [Concrete rule like "X > Y" or "always do Z"]
- [Timing/sequence: "if you do X before Y, Y fails"]
- [What to avoid: "don't default to [common mistake]"]
- [Meta-principle: iteration / measurement]
```
Why it works: Principles function as the prompt's "decision lens." The model uses them to weight competing considerations during generation. Contrarian principles ("perfectionism blocks need permission, not quality focus") produce non-obvious outputs.
Failure mode if absent: The output reverts to internet-median wisdom. Principles are how you steer the model away from default-mode thinking.
Frequency in our 545: 100%. The number of principles varies (3-7), but every Original has them.
Pattern 3: XML-tagged input fields
```xml
```
Why it works: XML tags are reliable structural delimiters across Claude, GPT, and Gemini. They make the prompt machine-parseable and force the user to think in the same fields the prompt will use.
Failure mode if absent: Prose-paragraph prompts produce variable output quality. Users skip fields. Models guess at missing context.
Frequency in our 545: 100%. No exceptions. (We tested without XML in early experiments. Quality dropped 25-40%.)
Pattern 4: Output-format that forces specific thinking
```xml
Diagnosis Title
First Required Section (often diagnostic)
[Forces structured analysis]
Second Section (often actionable framework)
[Forces specific recommendations]
Section requiring a table
[Forces ranked / comparable data]
Key Takeaways
[3-5 bolded bullets - non-negotiable]
```
Why it works: Generic output-formats ("Discussion") produce generic outputs. "Top 3 Failure Modes" produces failure-mode thinking. "Ranked by probability × test-cost" produces ranked thinking.
The trick: every section header forces a specific type of cognition.
Failure mode if absent: Output drifts. Models default to lists or paragraphs without internal structure.
Frequency in our 545: 100%. Section headers vary; the principle of forced-specific-thinking is universal.
Pattern 5: Auto-intake (the underused pattern)
```xml
If input incomplete: ask for [list of critical fields].
```
Why it works: Without this, users submit incomplete inputs and the model guesses (badly). With this, the model asks for the missing fields. Quality jumps 30-50% on incomplete inputs.
Failure mode if absent: Users provide partial context, the model fills gaps with assumptions, output is calibrated to invented context. The prompt fails silently.
Frequency in our 545: 100% in current Originals. We added it to 217 Originals in late 2025 after diagnostic data showed input-completeness was the largest variance source in output quality.
Pattern 6: The "Now, [verb]:" closing line
```
Now, run the diagnostic:
```
or
```
Now, build the framework:
```
Why it works: The closing imperative tells the model: "stop thinking, start producing." Without it, models sometimes generate meta-commentary about the prompt instead of executing it.
Failure mode if absent: Output begins with "I'd be happy to help you with..." or "Here's how I'd approach this..." — meta-talk that wastes tokens and signals uncertainty.
Frequency in our 545: 100%. Every Original ends with this pattern. It's a 1-line pattern that single-handedly improved output quality measurably.
Pattern 7: Definition sentence opening
The example output's first paragraph opens with:
A [type of thing] is a [definition]. Based on [source/statistic], [key fact about success/failure rates]. [Direct application to the specific case].
Why it works: This sentence pattern is what AI-Overview / Perplexity citation engines extract and quote. It's also the single most useful sentence for a reader skimming. AEO and human-readability align here.
Failure mode if absent: Outputs that open with "Looking at your situation..." or "Here are some thoughts..." — neither citable nor scannable.
Frequency in our 545: 100% in example outputs. We measured this directly in 2025: outputs opening with the definition sentence got cited 4.3× more often by AI Overviews than outputs opening with "Looking at..."
Section 3: What categories AI actually changes — and which it just decorates
This is the harder analysis. Across 545 Originals built and 8,500+ user feedbacks, here's what the data says about which categories AI is genuinely changing — versus which it's just adding latency to.
AI is genuinely changing (high signal, real value):
- Decisions & Reasoning — Five-scenario forecasters, pre-mortems, and steelmans produce thinking the user couldn't construct alone. AI's hybrid of breadth + neutrality is novel here. High retention in user data.
- Writing & Editing — Voice extraction, anti-bullshit grading, and structural diagnosis are genuinely useful. The user couldn't get this from a friend. High retention.
- Coding & Development — Codebase archaeology, code review with pattern-naming. AI sees the whole codebase faster than humans. High retention for senior engineers; mixed for juniors.
- Career & Work — Salary negotiation pre-mortems, boss-communication decoding, first-day diagnostics. The structured prep is real value. High retention during job-hunt phases, low otherwise.
- Memoir / Difficult-Life Writing — Vulnerability calibration, eulogy writing, grief processing. AI's neutrality is exactly what the user needs when their feelings are too close. Moderate retention, very high satisfaction when used.
AI is decorating (low signal, plausible-feeling output):
- Spiritual / Lifestyle — Generic "morning routine" or "manifestation" prompts produce plausible text that's indistinguishable from any wellness blog. Low retention in user data. We've stopped expanding this category.
- Generic productivity ("daily planning") — Outputs feel productive but rarely change behavior. The bottleneck is doing, not planning. Low retention.
- Most "AI-generated marketing copy" use cases — Output is competent and formulaic. Reads like every other AI-generated marketing copy. Effective when the underlying offer is strong, neutral when it's not. High volume usage, low repeat conversion.
The mid-tier (depends on user):
- Wellness & Health — Highly bounded prompts (burnout audit, energy audit, habit failure) work great. Generic ones (general "wellness check-in") don't. High variance.
- Relationships & Life — Conflict mediation works because the AI's neutrality is structural. Generic "relationship advice" doesn't. High variance.
- Money & Finance — Stress tests and red-flag inspectors work. Budget templates don't. High variance.
The pattern: AI helps where structure is the bottleneck and judgment is the input. AI decorates where the bottleneck is doing, not thinking.
Section 4: The 5 Highest-Engagement Originals (April 2026)
By daily-active-use across 545 Originals:
- Life Decision 5-Scenario Forecaster — Used heavily for major life-pivot moments (career changes, relocations, relationship decisions).
- Salary Negotiation Pre-Mortem — Spikes during compensation review cycles (March, September).
- Anti-Bullshit Essay Grader — Used by writers, PR professionals, students.
- Codebase Archaeologist — High concentration of senior-engineer users.
- Quit-or-Stay Calculator — Mid-career professionals make this query repeatedly across 6-12 months.
What unites them: each is for a specific high-stakes decision where the user's existing thinking is constrained by emotional proximity to the decision. AI's structural neutrality is the unlock.
What's not on the list: anything generic. The lowest-engagement Originals tend to be ones we built early when we hadn't fully diagnosed where AI helps versus decorates.
Section 5: What changed between 2024 and 2026
Three macro-shifts in how working prompts get built:
Shift 1: Markdown to XML
Two years ago, prompts often used markdown (# Section). Today, working prompts use XML (). Three reasons: better parseability across models, less ambiguity between content and structure, easier to nest hierarchically.
Cost of migration: re-writing 200+ existing Originals. Worth it: average output quality up ~15%.
Shift 2: Auto-intake became standard
In 2024, most prompts assumed the user would provide complete inputs. They didn't. Outputs were calibrated to invented context. Quality variance was huge.
In 2026, auto-intake is standard: if input is incomplete, the model asks for the missing fields before generating. Quality variance dropped significantly.
Shift 3: Example outputs became the contract
In 2024, prompts were sold/shared based on the prompt text alone. Quality was uncertain until you ran it.
In 2026, the example output IS the prompt's contract. If the example is bad, the prompt is bad. We make every Original ship with a 2,000-4,000 word example so users verify quality before use. This shifted Promptolis from "marketplace of prompts" to "verified library."
Section 6: The 5 patterns that fail in 2026 (avoid these)
Equally important — what we've stopped doing.
Anti-Pattern 1: "Comprehensive" prompts
Prompts that try to do everything. "Comprehensive career-and-life-and-finance audit." Result: generic output that's deep nowhere. Working prompts are narrow and specific.
Anti-Pattern 2: Voice-imitation
Asking AI to "write in my voice." Surface-level imitation works for 200 words; over 2,000 words it homogenizes. Working alternative: extract a voice spec sheet (see Style Archaeologist) and edit your own writing against it.
Anti-Pattern 3: "Just polish" prompts
"Make this better" applied to existing prose. Produces competent flatness. Detected by editors and AI-Overviews. Working alternative: structural audits + writer's own revision pass.
Anti-Pattern 4: Length-as-quality
Prompts that produce 5,000-word outputs are not better than ones that produce 1,500. They're harder to use. Optimal length: 1,500-3,000 words for complex Originals, 800-1,500 for medium, 500-1,000 for simple.
Anti-Pattern 5: Generic emotional prompts
"Help me with my anxiety." Outputs feel kind but produce internet-median wisdom. Working alternative: specific frameworks with named conditions and concrete next steps (see 3am Wake Rumination Interrupt for an example of how specificity beats generic empathy).
Section 7: What we'd build if we were starting over
If we were starting Promptolis from zero in 2026, knowing what we know:
- Start with 50 deeply-built Tier 3 Originals, not 500 broad ones. Volume comes later.
- Auto-intake as standard from day 1. It's the highest-leverage UX feature.
- Example outputs as the entry point, not the prompt text. Users decide on quality from outputs.
- Categorize by problem-type, not by topic. "Decisions you're avoiding" beats "Career & Work."
- Build the newsletter from launch. Email-list compounds; SEO doesn't fully replace it.
- Translate the top-50 Originals to 5+ languages early. ES + PT + DE + FR + IT covers ~1.5B speakers.
- Skip the marketplace temptation. Free + open is the durable wedge against PromptBase / AIPRM.
Section 8: What's coming next on Promptolis
By end of 2026:
- 750+ Originals across all 21 categories.
- Multilingual depth — DE / ES / PT / FR coverage of top 100 Originals.
- Interactive testers for each Original (paste input, get output, see structured analysis).
- API access for teams who want to embed Originals in their own workflows.
- Annual State of Prompts updates (this is the first; the next will analyze 750+ Originals).
If you're curious about specific Originals or analysis, the newsletter is where new ones land first.
Methodology Notes
- Dataset: 545 hand-crafted Promptolis Originals as of April 2026, plus 8,500+ aggregated user feedback signals from the prior 90 days.
- Patterns: Identified via inductive analysis. No formal NLP. Patterns are descriptive of working prompts, not statistical claims about all prompts.
- Engagement metrics: Daily-active-use (sessions where the user runs the prompt's example or pastes the prompt into an external tool, tracked via opt-in analytics). No personal data; no individual tracking.
- Limitations: We see what works on Promptolis users. They skew toward writers, founders, knowledge workers, and educators. Generalization to other audiences is informed but not guaranteed.
Cite this report
If you're writing about prompt engineering and want to cite this analysis:
Kürük, A. (2026). State of Prompts 2026: What 545 Hand-Crafted Originals Tell Us About AI Prompt Engineering. Promptolis. https://promptolis.com/blog/state-of-prompts-2026/
---