.webp "Advanced Prompt Engineering Techniques That Actually Work in 2026")
There's a before and after moment in prompt engineering that most practitioners can pinpoint precisely. For me, it happened after spending three hours wrestling a poorly structured AI output into something usable — only to realize the problem wasn't the model. It was my prompt.
That frustration sent me deep into the mechanics of how large language models actually process and respond to instructions. Over the past eighteen months of working with Claude, GPT-4, Gemini, and a handful of open-source alternatives, I've developed and refined a set of techniques that go well beyond surface-level prompt formatting. These aren't theoretical frameworks — they're approaches I apply daily on real projects, and they've meaningfully changed how I work.
What follows is a thorough breakdown of each technique, including the reasoning behind it, a ready-to-use prompt template, and an honest account of where it works best and where it falls short.
1. Recursive Self-Improvement Prompting (RSIP)
The single biggest shift in my output quality came from stopping the habit of accepting a model's first draft. Not because it's necessarily bad — modern LLMs produce surprisingly solid first passes — but because they have something most human editors lack: infinite patience for critique and revision. RSIP exploits this directly.
The idea is to structure your prompt so the model generates an initial output, systematically identifies its own weaknesses across specific dimensions, then improves on those weaknesses in a subsequent pass. Repeat this loop two or three times and you end up with something that has effectively been drafted, reviewed, and edited — all within a single prompt.
Why it works
Language models are trained on human feedback and extensive examples of revision. When you explicitly ask a model to critique its own work before improving it, you're activating a pattern the model has seen thousands of times in its training data: the edit-critique cycle that professional writers and engineers use. The model isn't just rewriting — it's identifying specific failure modes and directly addressing them.
The key detail most people miss: you must specify different evaluation criteria for each iteration. If you let the model pick what to fix, it tends to repeatedly address the most obvious surface-level issues — clarity, structure, tone — while ignoring deeper problems like logical gaps, unsupported assumptions, or missing edge cases. Forcing varied dimensions across iterations prevents this tunnel vision.
The prompt template
I need you to help me create [specific content]. Follow this exact process: Step 1: Generate an initial version of [content]. Step 2: Critically evaluate your own output against these criteria: - [Criterion 1: e.g., factual accuracy and specificity] - [Criterion 2: e.g., logical coherence and flow] - [Criterion 3: e.g., practical usefulness for the target reader] Identify at least 3 concrete weaknesses — be specific, not vague. Step 3: Create an improved version that directly addresses each identified weakness. Note what you changed and why. Step 4: Evaluate this improved version against a new set of criteria: - [Criterion 4: e.g., completeness — are there gaps or missing cases?] - [Criterion 5: e.g., conciseness — is anything redundant or unnecessary?] - [Criterion 6: e.g., audience alignment — does the tone match the reader?] Again, identify at least 3 specific weaknesses. Step 5: Produce a final, refined version addressing these new weaknesses. Step 6: Present only the final version, followed by a brief summary of the key improvements made across all iterations. Context about the target reader: [describe audience] Desired tone: [e.g., technical but accessible / conversational / authoritative]
Where it works best
RSIP has been most valuable for me in technical documentation, long-form content with complex arguments, and any situation where the first-draft quality gap between "acceptable" and "excellent" is wide. I've used it to produce API documentation that engineers actually read, strategy memos that executives find persuasive, and grant proposal sections that reviewers praised for clarity.
For short, factual outputs — a quick summary, a simple answer, a product description — the overhead isn't worth it. Save RSIP for work where quality genuinely matters more than speed.
Limitations to be aware of
RSIP increases token usage substantially — expect two to four times more tokens than a single-pass prompt. On models with strict context limits, long documents can run into issues if the full content plus critique plus revision all need to fit in one context window. For very long outputs, consider breaking the document into sections and running RSIP on each section independently.
Also worth noting: the improvement curve follows diminishing returns. The jump from iteration one to iteration two is usually dramatic. The jump from two to three is often marginal. I've stopped recommending more than three revision passes — the additional iterations rarely justify the cost.
2. Context-Aware Decomposition (CAD)
If you've ever asked an LLM to solve a genuinely complex problem — the kind with multiple interlocking constraints, dependencies that aren't obvious upfront, and tradeoffs that compound as you work through the solution — you've probably encountered what I think of as the "competent but shallow" failure mode. The model gives you something technically correct but structurally incomplete. It solves part of the problem while quietly ignoring other parts.
This happens because complex problems don't have obvious decomposition structures, and without explicit guidance, models tend to tackle whatever is most salient in the prompt rather than working systematically through the problem space. Context-Aware Decomposition addresses this directly by forcing the model to first map the problem before it solves it.
The difference from basic task splitting
Most people who use task decomposition simply break a big request into smaller requests. That helps, but it misses something crucial: the relationships between components. A business strategy problem where you separately analyze market position, operational efficiency, and team capacity will give you three disconnected analyses. CAD forces the model to explicitly reason about how each component affects the others — and to carry that awareness through the synthesis step at the end.
The prompt template
I need to solve the following complex problem: [describe problem in full detail] Please work through this using the following structured approach: Phase 1 — Problem Mapping Identify the 3–5 core components of this problem. For each component: a. Name it clearly and explain why it matters to the overall solution b. Identify what information or reasoning is needed to address it c. Flag any dependencies — does solving this component require solving another first? Phase 2 — Component-Level Analysis Work through each component identified in Phase 1 separately. For each: a. Develop a focused solution or analysis b. Note any assumptions you're making and why c. Flag any interactions with other components that will need to be addressed in synthesis Phase 3 — Synthesis Combine the component-level analyses into a coherent, integrated solution. Explicitly address: a. How the components interact or create tradeoffs b. Any tensions or conflicts between component-level solutions c. What you had to prioritize or deprioritize, and why Phase 4 — Final Recommendation Present a clear, actionable answer or recommendation that accounts for the full complexity of the problem. Include the top 2–3 risks or caveats. Maintain a running "dependency log" throughout — whenever a decision in one component affects another, note it explicitly.
Real application: systems architecture
One of the clearest demonstrations of CAD's value came when I was using it for a systems architecture review. A basic prompt produced a reasonable architecture but completely missed the tension between the latency requirements of the real-time processing layer and the cost constraints in the data storage design. Running the same problem through CAD produced an architecture that explicitly called out this tension in Phase 2, then resolved it in synthesis with a tiered caching approach neither I nor the initial prompt had considered.
The model didn't become smarter. It became more disciplined — which is often exactly what's needed for complex problems.
Best use cases
- Systems architecture and technical design problems with multiple constraints
- Business strategy decisions that span multiple functions (marketing, operations, finance)
- Research synthesis where multiple bodies of evidence need to be integrated
- Complex debugging where multiple interacting issues might be present
- Policy analysis where second-order effects matter as much as first-order outcomes
When to skip it
CAD is genuinely overkill for well-defined problems with a clear single answer. If the path from question to solution is relatively linear, the decomposition overhead adds length without adding insight. Use it when the problem has genuine structural complexity — multiple constraints, interdependencies, or competing priorities — not simply when the problem feels hard.
3. Controlled Hallucination for Ideation (CHI)
This one tends to raise eyebrows, so let me be direct about what it is and what it isn't. Hallucination — the tendency of language models to generate confident-sounding but factually incorrect content — is a genuine problem when you need accurate information. But it's a surprisingly useful property when you need genuinely novel ideas.
The model's ability to generate plausible-sounding things that don't yet exist is, in a controlled context, exactly what brainstorming and speculative ideation require. CHI is a structured technique for deliberately activating this tendency in a bounded, clearly labeled way to generate creative concepts that sit at the edge of what's currently possible.
The underlying logic
Language models are extraordinarily good at pattern recognition across the vast landscape of human knowledge. When you ask for genuinely novel ideas, the model extrapolates from patterns it has seen — connecting concepts from different domains in ways that humans might not naturally consider because we're subject to cognitive biases and disciplinary silos. CHI harnesses this extrapolation deliberately, then uses a feasibility filter to separate the genuinely interesting ideas from the purely fantastical.
In practice, about 25–35% of CHI-generated speculative ideas survive a rigorous feasibility filter as concepts worth exploring further. That's a significantly higher yield than most conventional brainstorming approaches, where the constraint of "must exist" or "must be obviously feasible" dramatically limits the idea space.
The prompt template
I'm working on [specific domain/problem/project] and need genuinely innovative ideas — including concepts that don't currently exist but could. Please engage in structured speculative ideation: Step 1 — Speculative Generation Generate 6–8 hypothetical innovations or approaches for this domain. These should be genuinely novel — not existing solutions with different names. For each one: a. Give it a clear, specific name and one-sentence description b. Explain the theoretical mechanism or principle that would make it work c. Describe what it would look like in practice — what would a user actually experience? d. Identify the key technical or practical barriers to implementation today Label every idea clearly as [SPECULATIVE — NOT CURRENTLY AVAILABLE]. Step 2 — Feasibility Triage Review all 8 ideas and assess each against two dimensions: - Technical readiness: How close are we to having the underlying capabilities? - Impact potential: If it worked, how significant would the benefit be? Create a 2x2 matrix in text form: High Feasibility / High Impact (pursue), High Impact / Low Feasibility (long-term watch), Low Impact / High Feasibility (low priority), Low Impact / Low Feasibility (discard). Step 3 — Deep Dive Take the 2–3 ideas from the "pursue" quadrant and develop each one further: a. What would a minimum viable version of this look like? b. What existing technologies or approaches could be repurposed to build toward it? c. What would need to be true — technically, economically, or socially — for it to reach viability? Throughout: maintain clear labeling. Do not present speculative ideas as existing solutions.
Where I've used this successfully
CHI has been most valuable in three areas: product strategy sessions where I need to get past "what are competitors doing" thinking, research direction brainstorming where I want to explore the edges of a field, and creative projects where novelty is the primary success criterion.
One specific example: working on a content personalization problem, CHI generated a concept for what it called "narrative-state-aware recommendation" — a speculative approach that treated content consumption as a story with a reader's current emotional and informational state as the recommendation input, not just their click history. It didn't exist as described. But the concept was interesting enough that a deeper investigation revealed two academic papers exploring adjacent ideas that the team hadn't encountered. The speculative concept became a useful conceptual frame for the actual work.
Practical limitations
CHI outputs are only as good as your domain framing. If you're vague about the problem space, the model generates vague speculative ideas. The more specific and bounded the domain you provide, the more targeted and actionable the speculative outputs tend to be. Also, some models are more cautious than others about generating speculative content — Claude and GPT-4 respond well to this technique; some smaller models will default to describing existing solutions instead of generating genuinely novel concepts.
4. Multi-Perspective Simulation (MPS)
Most AI-generated analysis suffers from a subtle but significant flaw: it arrives at a conclusion too quickly. Present a complex question and you get a competent, organized answer — but one that tends to reflect whatever viewpoint was most represented in the model's training data for that topic, presented with more confidence than the underlying complexity warrants.
Multi-Perspective Simulation is a technique for forcing the model to inhabit multiple distinct viewpoints on a question before synthesizing them. The result isn't just a more balanced analysis — it's a qualitatively different kind of thinking that surfaces considerations you wouldn't have reached through a standard analytical prompt.
The key design principle
The most common mistake with this technique is defaulting to simplistic pro/con or agree/disagree structures. "Here are three arguments for and three arguments against" isn't perspective-taking — it's list-making. True MPS requires the model to inhabit each perspective with genuine depth: understanding its underlying values and assumptions, articulating its strongest arguments (not the strawman version), and honestly acknowledging its real weaknesses.
The "intellectual charity" instruction in the template below is not decorative. It directly affects output quality by preventing the model from treating any perspective as obviously inferior — which in turn forces more rigorous engagement with each viewpoint's actual reasoning.
The prompt template
I need a thorough, intellectually honest analysis of [topic/question/decision]. Please create a multi-perspective simulation using this structure: Step 1 — Perspective Identification Identify 4–5 meaningfully distinct perspectives on this issue. These should differ in their underlying values, frameworks, or epistemic approaches — not just their surface-level conclusions. Briefly describe each perspective's core worldview and who holds it in practice. Step 2 — Deep Perspective Analysis For each perspective: a. State its core assumptions and values — what does this perspective believe about human nature, institutions, tradeoffs, etc.? b. Present its strongest version of the argument — the case a thoughtful, well-informed proponent would make c. Identify its genuine blind spots, empirical weaknesses, or internal tensions — be specific, not dismissive Maintain intellectual charity throughout. Every perspective should receive the same quality of attention. Step 3 — Structured Dialogue Simulate a constructive dialogue between two or three of the most opposed perspectives. Show: a. Where they talk past each other (different definitions of success, different empirical assumptions) b. Where they might actually agree if they shared common premises c. What evidence or argument would be most likely to shift each perspective Step 4 — Synthesis Provide an integrated analysis that: a. Acknowledges what each perspective gets right b. Identifies the crux questions — the empirical or values disagreements that most drive the differences c. Offers a nuanced conclusion that doesn't collapse the complexity Do not present the synthesis as "the truth" — present it as the most defensible position given current evidence and explicit acknowledgment of remaining uncertainty.
When this changes the outcome
I've used MPS most extensively for policy analysis and complex organizational decisions. The consistent pattern: the synthesis step routinely surfaces considerations that weren't in my original framing of the question. A technology adoption analysis that I framed as a capability vs. cost question turned into a much richer analysis that included a "change management" perspective and an "institutional risk" perspective I hadn't originally considered. The final recommendation was different — and better — because of it.
For personal decisions or moderately complex questions with relatively clear answers, MPS adds overhead without proportional value. Reserve it for questions where you genuinely believe multiple serious, legitimate perspectives exist.
A note on political and socially contentious topics
MPS works well for policy and ethical questions, but be intentional about perspective selection. If you're not careful, the model can default to a left/right or progressive/conservative frame that doesn't actually capture the most important distinctions. For any complex topic, explicitly name the perspectives you want explored rather than letting the model choose — and push for perspectives that cut across conventional political lines when relevant.
5. Calibrated Confidence Prompting (CCP)
Of all the techniques in this article, this is the one I wish had been developed earlier. The problem it solves is one of the most persistent and genuinely dangerous failure modes of language models: the presentation of uncertain or incorrect information with the same confident, authoritative tone as well-established facts.
A model that says "The study found X" and "It's widely believed that Y" with identical confidence levels — regardless of whether X is from a rigorous meta-analysis or a single blog post, and regardless of whether Y is a scientific consensus or a fringe hypothesis — is actively misleading. Not because it's lying, but because the confidence signal is absent.
Calibrated Confidence Prompting adds an explicit, structured confidence layer to every claim the model makes.
Why standard outputs are poorly calibrated
Language models are trained to produce fluent, coherent text. Fluency tends to correlate with confidence — hedged, uncertain language often sounds less well-written than declarative statements. This creates a systematic pressure toward overconfidence in outputs. Without explicit prompting to calibrate confidence, models will default to sounding more certain than the underlying evidence warrants.
This isn't a model flaw in the traditional sense — it's a feature of the training objective that becomes a problem in high-stakes applications. CCP is the practical workaround.
The confidence scale
The scale below has been refined through about eight months of use. The percentages are approximate and meant to anchor intuition, not to imply false precision about model confidence levels.
Virtually Certain (>95%): Well-established facts, mathematical truths, physical constants. No meaningful doubt. Highly Confident (80–95%): Strong evidence from multiple independent sources. Real exceptions may exist but are uncommon. Moderately Confident (60–80%): Reasonable evidence, but meaningful uncertainty remains. Conflicting studies or expert disagreement exists. Speculative (40–60%): Reasonable conjecture based on limited evidence or extrapolation from adjacent domains. Should be treated as a working hypothesis. Unknown / Cannot Determine: Insufficient evidence to form even a speculative view. Model should flag this explicitly rather than guessing.
The full prompt template
I need information about [topic]. This will be used for [describe application — e.g., research, a decision, content creation]. When responding, apply explicit confidence calibration to your claims: For each substantive claim or statement, indicate its confidence level using: [VC] = Virtually Certain (>95%) [HC] = Highly Confident (80–95%) [MC] = Moderately Confident (60–80%) [SP] = Speculative (40–60%) [UK] = Unknown / Cannot Determine Additional requirements: 1. For [VC] and [HC] claims: briefly name the basis (e.g., "per WHO guidelines," "consistent finding across 15+ studies") 2. For [MC] and [SP] claims: name the key source of uncertainty and what evidence would resolve it 3. For any claim you're tempted to make confidently but aren't certain about: default to a lower confidence rating 4. If asked about my specific domain context [add details], flag any claims that are well-established in general but might not hold in my specific context Prioritize accurate calibration over sounding authoritative. I'd rather know what you don't know than receive confident misinformation.
Where this has the most impact
CCP has transformed how I use LLMs for research, due diligence, and anything where decisions are based on the output. The confidence labels do two things simultaneously: they make the output more honest, and they make it more useful, because I can immediately see which claims need independent verification and which I can reasonably rely on.
In one particularly useful application, I ran a competitive analysis through CCP prompting and discovered that about 40% of what I would normally have treated as solid market information was actually rated Speculative or Moderately Confident by the model. That changed how I weighted those inputs and prompted additional research in exactly the areas where uncertainty was highest.
Calibrating the calibration
One practical note: models vary significantly in how well they implement confidence calibration. Some are prone to assigning [HC] to almost everything — a form of systematic overconfidence within the framework itself. If you notice this pattern, add an explicit check: "Review your confidence ratings after completing your response. If more than 60% of your substantive claims are rated [HC] or [VC], you are likely over-calibrating. Revise the ratings that are most questionable."
6. Combining Techniques for Maximum Impact
Each technique in this article is valuable on its own. But some of the most powerful prompting work I've done involves combining two or three of them within a single structured prompt. The combinations aren't arbitrary — each pairing addresses a different dimension of output quality simultaneously.
RSIP + CAD: For complex, high-quality deliverables
Use CAD's decomposition structure to map and solve a complex problem, then apply RSIP's iterative critique-and-improve loop to the final synthesis. This combination produces outputs that are both structurally complete (CAD's contribution) and polished (RSIP's contribution). I use it for strategy documents, technical proposals, and long-form analysis where both depth and quality matter.
MPS + CCP: For research and analysis under uncertainty
Run a multi-perspective simulation to ensure comprehensive coverage of a topic, then apply confidence calibration to the synthesis. The result is an analysis that's both intellectually balanced and epistemically honest — you get the nuance of multiple viewpoints plus clear signals about which claims are well-supported versus speculative. Particularly valuable for literature reviews and evidence summaries.
CHI + RSIP: For creative and innovation work
Use CHI to generate a pool of novel concepts, then apply RSIP's iterative refinement to the most promising ideas. This combination helps you move from speculative ideation to genuinely developed concepts without losing the creative ambition that CHI generates. The critique iterations tend to strengthen the most viable ideas and expose fatal flaws in the less viable ones.
CAD + CCP: For decisions with significant uncertainty
For complex decisions where some information is solid and other information is uncertain, pair CAD's systematic component analysis with CCP's confidence calibration. The result is a structured analysis where you know not just what the components are, but how confident you should be in each component's analysis. This is the combination I use for investment-related analysis and significant business decisions.
7. Real-World Results and Benchmarks
Numbers can be slippery when applied to qualitative improvements, but across eighteen months of consistent application, these are the patterns I've observed across projects:
One important caveat: these numbers reflect patterns across my specific use cases and working style. The techniques perform significantly better when the underlying prompt is well-constructed — sloppy inputs still produce sloppy outputs, regardless of the framework applied on top. These techniques are multipliers, not substitutes for clear thinking about what you actually need.
8. Common Mistakes and How to Avoid Them
After teaching these techniques to colleagues and watching where they struggle, the same patterns come up repeatedly.
Mistake 1: Applying advanced techniques to simple problems
RSIP for a two-sentence summary. CAD for a straightforward question with an obvious answer. MPS for a question where one perspective is clearly correct. Technique selection should match problem complexity — overthinking simple tasks produces longer outputs, not better ones. A quick heuristic: if you could solve the problem with a single clear sentence of instruction, you probably don't need these techniques.
Mistake 2: Treating the templates as fixed
The templates in this article are starting points, not scripts. Every effective prompt I use in practice has been adapted for the specific task, audience, and output format. The underlying structure of each technique — the logic of iterative critique, systematic decomposition, controlled speculation, perspective-taking, and confidence calibration — is what matters. The specific wording should evolve as you understand your use case better.
Mistake 3: Skipping the evaluation dimensions in RSIP
The most common failure mode in RSIP is letting the model choose its own evaluation criteria without guidance. Without specified dimensions, models consistently default to surface-level fixes — rephrasing for clarity, improving transitions, adding examples. These are fine improvements, but they miss deeper issues. Always specify what dimensions you want evaluated, and vary them across iterations.
Mistake 4: Not anchoring CHI with enough domain specificity
Vague CHI prompts produce vague speculative ideas. "Generate innovative ideas for healthcare" produces much weaker output than "Generate innovative approaches to reducing medication non-adherence in elderly patients managing multiple chronic conditions." The more precisely you define the problem space, the more targeted and actionable the speculation becomes.
Mistake 5: Using MPS as a debate generator rather than an analysis tool
MPS is not a debate exercise. The goal isn't to find the "winner" among perspectives — it's to understand the problem more completely by seeing it through multiple legitimate lenses. If your synthesis is declaring one perspective correct and dismissing the others, you've turned MPS into a more elaborate version of confirmation bias. The synthesis should genuinely integrate insights from all perspectives, even perspectives you personally find less compelling.
Mistake 6: Ignoring low confidence ratings in CCP outputs
The whole point of CCP is to surface uncertainty so you can act on it. If you note a cluster of [SP] and [MC] ratings in a section of the output and proceed as though they were [VC], you've done the work of calibrating without getting the benefit. Treat low-confidence ratings as action items: these are the claims that need independent verification before you rely on them.
9. What Comes Next in Prompt Engineering
The most honest thing I can say about the future of this field is that its half-life is shortening. Techniques that were cutting-edge eighteen months ago are now table stakes. The five techniques in this article are the best approaches I know today — but "today" is doing a lot of work in that sentence.
A few directions I'm actively exploring, and where I think the most interesting work is happening:
Meta-prompting and adaptive instruction sets
Instead of writing a fixed prompt for a specific task, designing prompts that can adapt their approach based on what the model encounters as it works through a problem. Early experiments suggest this can dramatically improve performance on problems where the right decomposition strategy isn't obvious upfront.
Structured uncertainty propagation
An evolution of CCP that doesn't just label uncertainty at the level of individual claims, but tracks how uncertainty compounds across a chain of reasoning. When conclusion C depends on claim B which depends on claim A — and A is only Moderately Confident — the uncertainty doesn't just add, it multiplies. Prompting models to reason explicitly about uncertainty propagation is technically harder but yields significantly more calibrated outputs on complex analytical tasks.
Constraint-first prompting
Rather than describing what you want, leading with what you absolutely cannot accept. This mirrors how engineers think about system design — constraints are more definitive than goals — and I've found it produces meaningfully different outputs on problems where failure modes matter more than optimizing for the ideal case.
Multi-agent collaboration frameworks
Using multiple model calls in structured roles — one model generating, another critiquing, a third synthesizing — rather than trying to get a single model to do all three simultaneously. This is more expensive and architecturally complex, but on high-value tasks the quality improvement can be substantial. As model costs continue to decline, this approach will become more accessible for everyday use.
The most fundamental shift in prompt engineering over the next few years won't be new techniques — it will be a clearer understanding of what prompts actually do inside these systems. As interpretability research matures, we'll move from empirical pattern-finding to principled design.
For now, the techniques above represent the most reliable tools I have. They work because they align with how capable language models actually process complex instructions — not because they involve any magic or model-specific tricks. That's both their limitation and their durability: as models change, the underlying principles should continue to apply.
Final Thoughts
Prompt engineering sometimes gets dismissed as a temporary skill — something that will become unnecessary as models get smarter. I've never found that argument convincing, and I find it less convincing now than I did a year ago. The techniques in this article aren't workarounds for model limitations. They're structured thinking frameworks that help you get the most out of systems that are genuinely capable of sophisticated reasoning — if you know how to engage them properly.
The most consistent finding across all five techniques: explicit structure in your prompts produces more reliable, higher-quality outputs than hoping the model infers what you need from a well-phrased request. Models have gotten remarkably good at inference. They haven't gotten so good that explicit structure is unnecessary.
If you apply only one technique from this article, make it Calibrated Confidence Prompting. The habit of demanding explicit confidence signals from AI outputs will improve the quality of every decision you make based on that information — regardless of what other prompting approaches you use.
If you're working on genuinely complex problems and have more time to invest, the CAD + CCP combination offers the best return on effort. You get structural rigor and epistemic honesty simultaneously, which covers the two most common failure modes of standard AI-assisted analysis in one framework.
Start with a single technique on a real project — not a test case, but something that actually matters. The learning from one applied use case is worth more than reading about all five techniques without trying any of them.