A small observation about AI image generation can open into something much larger. People are getting better results from image models when they stop giving vague wishes and start giving precise briefs. That sounds like a practical note about prompting. But beneath it is a deeper principle: intelligence, whether biological, human, or artificial, works better when intention is converted into structure.
The phrase that started this line of thought was directed specificity. Not mere detail. Not a longer prompt. Not a blizzard of adjectives. Directed specificity: the ability to tell a system which part of possibility-space matters, what is fixed, what is flexible, what counts as success, and what would count as failure.
That principle is useful for image generation. But it may also point toward something much more important: a way of understanding what Clean Well is actually for.
A precise instruction does not make a claim true. A detailed brief does not prove anything. Specificity does something different: it directs attention. It narrows the field. It tells the system where to look and how to judge what it finds.
That is why vague prompts produce generic outputs. If the user says, “make it beautiful,” the model has to guess what beautiful means. If the user says, “make it premium, restrained, commercially publishable, with a clean typographic hierarchy and no cheesy sci-fi imagery,” the model has less guessing to do. The work has been constrained.
But there is a catch. More detail is not always better. Detail can become noise. A prompt can be long and still confused. It can contain contradictory instructions, undefined aesthetic terms, hidden priorities, and emotional pressure disguised as specification. So the real skill is not adding detail. The real skill is finding the details that matter.
From Prompting to Clean Well
This is where Clean Well enters. Clean Well should not be thought of as merely a better prompt format. That would undersell it. Nor should its categories simply be pasted into prompts as decorative headings. That would be the shallow version of the idea.
The stronger possibility is that Clean Well is a prompt-construction layer. It does not just write the prompt. It works out what the prompt needs to know.
Most bad prompts are not bad because the user is unintelligent. They are bad because human intention arrives compressed. We know roughly what we mean, but the different layers are tangled together. We mix what something is with how it should work, what we know with what we prefer, what must be preserved with what can be changed, what is factual with what is aesthetic.
Clean Well’s value is that it separates those layers before the task begins.
Messy human intention
↓
Clean Well decomposition
↓
Typed dependency structure
↓
Engineered prompt / brief / specification
↓
Better downstream output
In this framing, Clean Well is not just a way to ask AI for better images. It is a way to convert vague intention into executable structure.
The Five Kinds of Constraint
One way Clean Well can do this is by separating a language packet into different kinds of dependency. In the current schema, these include definitional, empirical, mechanistic, normative, and procedural elements.
| Type | Question it answers | Prompt function |
|---|---|---|
| Definitional | What is this thing? What do the key terms mean? | Defines the task, object, role, boundaries, and success criteria. |
| Empirical | What facts, measurements, sources, or constraints are fixed? | Grounds the task in real data, dimensions, names, dates, files, or specifications. |
| Mechanistic | How is this supposed to work? | Clarifies internal structure, hierarchy, causality, composition, and functional logic. |
| Normative | What values, preferences, standards, or judgements apply? | Specifies taste, tone, desired quality, and what should be avoided. |
| Procedural | What steps should be followed? | Controls workflow, order of operations, preservation rules, and output format. |
Consider a book-cover prompt. A weak version might say: “Make this more modern and professional.” That sounds meaningful to the person saying it, but it leaves almost everything important unstated.
Clean Well would ask: What does “modern” mean here? What must not change? What exact file dimensions matter? Is the desired style commercial nonfiction, academic seriousness, literary prestige, or popular science? Should the typography dominate, or the image? Is the instruction aesthetic, procedural, empirical, or all three?
The result is not necessarily a longer prompt. It is a cleaner one.
Prompt Hygiene
This leads to one of the more practical uses of Clean Well: prompt hygiene.
Many prompts are polluted without appearing obviously wrong. They contain contradictions, vague adjectives, competing goals, undefined standards, repeated instructions, and hidden assumptions. They may feel rich to the user, but to the model they are a bundle of partially incompatible pressures.
“Make it bold, elegant, minimal, dramatic, subtle, timeless, futuristic, warm, serious, and eye-catching” is not a brief. It is an aesthetic traffic jam.
Clean Well’s contamination flags could identify these problems before execution. It could point out where language is smuggling in a value judgement, where two instructions conflict, where a term shifts meaning, or where rhetorical force is substituting for actual constraint.
This is where the system becomes more than taxonomy. It becomes a calibration instrument.
The Drill Target
Perhaps the most important feature is the single best drill target. The system should not ask for endless clarification. That would make it annoying and inefficient. Instead, it should identify the clarification with the highest expected leverage.
In a design brief, the crucial missing point might be whether the work should feel academic or commercial. In an argument, it might be an undefined term. In a legal email, it might be the difference between a factual statement and a procedural requirement. In a manuscript, it might be whether a claim is being offered as metaphor, mechanism, or evidence.
The drill target is the place where the next unit of attention buys the most clarity.
This phrase matters. It protects the system from becoming a verbose ontology machine. The aim is not to decompose everything forever. The aim is to find the structure required to make the next move better.
From Better Prompts to Visible Reasoning
At this point, the idea widens. If Clean Well can decompose a prompt before generating an artefact, perhaps it can also decompose a claim before accepting or rejecting it.
A claim often arrives as a single smooth sentence: “This is a good idea.” “AI will transform work.” “This argument is weak.” “This design is better.” But each of those sentences compresses a whole dependency structure.
What does “good” mean? Good for whom? Under what constraints? What empirical facts would need to be true? What mechanism is supposed to produce the benefit? What values are being applied? What evidence is missing? What ambiguity is being hidden by the smoothness of the sentence?
Clean Well can unfold the sentence into its load-bearing parts.
Expressed claim ├── Definition: what key terms mean ├── Empirical dependency: what would need to be true ├── Mechanistic dependency: how the thing is supposed to work ├── Normative dependency: what value or standard is being applied ├── Procedural dependency: what method or step is assumed ├── Ambiguity: what remains unresolved ├── Contamination: where language is doing hidden work └── Drill target: the next highest-leverage clarification
This is why Clean Well begins to look like a reasoning microscope. It does not reveal raw thought. That would be too strong. Human cognition includes memory, emotion, bodily state, unconscious prior, social context, and association. Clean Well cannot see all of that.
What it can do is reconstruct the publicly inspectable dependency structure of language once thought has been expressed.
The Compiler Thesis
This leads to the central architectural insight: Clean Well is not mainly a prompt-engineering trick and not mainly an argument-analysis tool. It is part of a compiler architecture.
Human language is compressed source material. Clean Well parses it. The typed dependency graph is the intermediate representation. Then that graph can be rendered into whatever the task requires: a prompt, a critique, a rewrite, an evidence plan, a disagreement diagnosis, or a test matrix.
Compressed language packet
↓
Clean Well parser / type-checker
↓
Typed dependency graph
↓
Target rendering
├── prompt
├── brief
├── critique
├── rewrite
├── evidence plan
├── argument map
└── disagreement diagnosis
This matters because it explains why the same system seems relevant in different places. Image prompts, manuscript edits, philosophical claims, property emails, product specifications, and research hypotheses are not the same kind of thing. But they are all language-mediated tasks where ambiguity carries cost.
The typed dependency graph is the centre. Everything else is a rendering.
Clean Well and CRF
There is also an important architectural distinction. Clean Well should not be overloaded as the whole system. Clean Well is better understood as the parser and type-checker. The wider CRF system is the compiler.
The decomposition organ
Takes messy language and breaks it into typed dependencies, contamination risks, ambiguities, and drill targets.
The compiler architecture
Uses the typed graph to generate downstream artefacts: prompts, briefs, critiques, rewrites, evidence plans, and tests.
Other organs then become clearer. Sentinel audits calibration, overreach, and contamination handling. Reality Bridge connects unresolved empirical dependencies to external verification. The synthesis layer recompiles the graph into useful outputs.
Once this distinction is made, the system becomes easier to explain and harder to overclaim.
Forward Mode and Backward Mode
The same machinery works in two directions.
In forward mode, Clean Well starts with an intention and turns it into a specification. This is the brief-building side. A vague wish becomes executable.
In backward mode, Clean Well starts with a claim and exposes its dependency skeleton. This is the reasoning-inspection side. A smooth sentence becomes an argument map.
Forward mode: intention → typed graph → specification Backward mode: claim → typed graph → diagnosis
This is the important part: these are not two unrelated methods. They are two uses of the same middle structure. One looks forward toward action. The other looks backward toward justification.
That is why the system has power. It can help us make things, and it can help us understand what our claims depend on.
The Ontology Is an Engineered Carving
There is a risk here. The five Clean Well categories can start to look like natural kinds: as if reality itself were divided cleanly into definitional, empirical, mechanistic, normative, and procedural elements.
That would be a mistake.
These categories are better understood as an engineered carving. They are not a final metaphysics of thought. They are a practical way of separating language into useful parts.
In reality, the categories overlap. Mechanistic claims often contain empirical assumptions. Procedural claims often smuggle normative priorities. Definitions can encode values. Empirical claims depend on measurement procedures.
This is not a flaw if it is handled honestly. It is exactly why contamination flags matter.
What Is Original Here?
The five categories alone are probably not the most original part. Many traditions already separate facts, values, definitions, mechanisms, and methods in one form or another.
The originality sits elsewhere.
It sits in the combination of typed decomposition, contamination detection, ambiguity preservation, single best drill target, and downstream recompilation. That combination turns Clean Well from a taxonomy into a working instrument.
The contamination flags stop the language from pretending to be cleaner than it is. The drill target prevents the system from becoming bloated. The typed graph preserves the structure. The synthesis layer turns the analysis back into action.
This is the difference between a list of categories and a genuine cognitive tool.
Can This Become Science?
There is a serious scientific paper hiding here, but only if the claim is tightened.
The weak version says: “Clean Well reveals the structure of thought.” That is too large, too philosophical, and too hard to falsify.
The stronger version says: “Clean Well is a typed dependency-graph framework for decomposing natural-language claims and instructions into inspectable structures, and those structures improve downstream reasoning, prompting, and specification tasks.”
That can be tested.
Image generation could be a first benchmark because the failures are visible. Take messy prompts, compare raw prompts against Clean-Well-compiled prompts, and measure adherence, preference, revision count, and error severity.
But the broader claim is not limited to images. The same test could apply to manuscript editing, argument analysis, product specifications, legal communications, AI-agent instructions, and disagreement diagnosis.
The first test should be small. Take one real artefact and run the system both ways. Decompose it backward into a dependency graph. Then recompile that graph forward into a better output. If the same graph supports both understanding and execution, the compiler thesis gets stronger.
The Deeper Aside: Life as Argument
At this point, the idea widens beyond language.
Perhaps life is always making an argument. Not a verbal argument. Not a conscious syllogism. But an embodied wager.
Every living system acts before certainty is available. It senses, compares, commits energy, moves, tests, and updates. A bacterium moving up a nutrient gradient is not thinking in language, but its structure encodes a kind of claim: this signal has mattered before; movement in this direction tends to preserve viability; continue.
A fin is an argument about water. A wing is an argument about air. A root system is an argument about soil, gravity, water, and minerals. An eye is an argument that light contains exploitable structure. A nervous system is an argument that internal modelling is worth the metabolic cost.
The difference across life is scale and complexity. Bacteria make shallow chemical wagers. Animals make sensorimotor wagers. Humans make recursive symbolic arguments. Culture and science then externalise those arguments into shared systems.
This links Clean Well back to the larger project. Life bets on patterns. Language compresses those bets into claims and intentions. Clean Well decompresses those claims and intentions so their structure can be inspected, repaired, tested, and recompiled.
The Shape of the Whole Journey
The path looks like this:
AI image briefs
↓
Directed specificity
↓
Specificity as search-space control
↓
Clean Well as prompt-construction layer
↓
Typed dependency decomposition
↓
Contamination flags and drill targets
↓
Reasoning visibility
↓
Typed dependency graph as intermediate representation
↓
CRF as compiler architecture
↓
Scientific testing
↓
Life as embodied argument under uncertainty
The hidden continuity is this: language arrives compressed. It hides dependencies. It fuses fact, value, mechanism, method, and preference into smooth sentences. Clean Well is an attempt to decompress that language into visible structure.
Once visible, the structure can be inspected. Once inspected, it can be corrected. Once corrected, it can be recompiled into better action.
The Current Best Formulation
The best current formulation is probably this:
A shorter version would be:
And the most compressed version might be:
Why This Matters
We are entering a world where models can generate images, essays, code, designs, plans, arguments, and decisions from language. In that world, the bottleneck shifts. The limiting factor is no longer only the model’s ability to execute. It is the human ability to specify what execution should mean.
Prompt engineering, in the old sense, was often treated as a bag of tricks. But the deeper skill is not tricking the model. It is clarifying the task. It is converting intention into structure. It is knowing what kind of specificity matters.
That is why Clean Well matters. It is not just a way to get prettier images or cleaner outputs. It is a way of making language less foggy. It gives us a method for asking: what is being claimed, what is being assumed, what is being valued, what is being hidden, and what must be clarified next?
In a world increasingly mediated by language models, that is not a minor skill. It may become one of the central cognitive disciplines.
Life acts by betting on patterns. Human language compresses those bets into claims, stories, arguments, designs, and instructions. Clean Well is an attempt to unfold those compressed packets, reveal their hidden dependencies, and turn blurred intention back into structured action.
That is the real journey: from better prompts to visible reasoning; from visible reasoning to typed dependency graphs; from typed dependency graphs to a possible science of language-mediated action.