The Art of the Prompt
Brief an LLM so it reads your creative work with discipline, not flattery
The problem with most AI feedback is its eagerness to please. Ask a large language model (LLM) if a scene is working, and it will likely offer vague praise, hedging its bets to keep the user happy. This flattery is useless for any writer serious about their craft.
To get sharp, analytical feedback, you have to brief an LLM with the precision of an engineer and the nuance of a writer. You must replace the vague request with a structured set of roles, constraints, and contracts. In my prompts for creative writing analysis, this often comes together in five components:
Role – the stance it should adopt.
Scope – the boundaries of its reading.
Procedure – the steps it must follow.
Rubric – the standards it should apply.
Output Contract – the format it delivers.
Each component shapes the model’s response in the way that voice, structure, and theme shape a piece of creative writing or the way specifications, schematics, and tolerances shape an engineered system. The closer the elements work together, the clearer, more consistent, and more purposeful the result.
A Practical Demonstration
Look at how a well-crafted prompt (available below) behaves with a scene fragment that has atmosphere but lacks dramatic propulsion.
Input scene (fragment):
Mara set two plates on the table though only one chair was pulled out. Steam curled from the pot on the stove. She poured herself a glass of water, took one sip, then left it half-full. She glanced at the clock, stood by the window, and pressed her forehead against the glass.
Output excerpt:
==RUBRIC-EVALUATION==
A) Absent: No explicit external goal; actions suggest waiting but lack a defined in-scene objective.
B) Absent: No clear irreversible turn; the sequence builds tension without a pivotal, binding shift.
C) Partial: Sensory details like "steam curled" evoke atmosphere, but they're static, not tied to character action.
D) Strong: Unspoken longing or abandonment hums beneath the domestic rituals.
E) Strong: Diction avoids clichés, opting for precise, evocative phrasing.
==REVISION-STRATEGY==
To inject an irreversible turn, identify a mid-scene threshold (e.g., the clock hitting a deadline) that forces Mara into a concrete action, like dialing a number or scraping a plate, binding her solitude without resolving the wait.
Notice how the engine praises the diction and subtext while flagging the structural weakness - a lack of goal and turn. The feedback is balanced, analytical, and actionable. This short demo shows the value of structure: once you set the criteria, the model reads to them.
The Analysis Engine: A Core Prompt
The Analysis Engine below is a baseline prompt for editorial work on creative writing. It’s a tool for diagnosis. It won’t replace your taste or voice. Think of it as a professional reader. It was created with fiction in mind but it would also work with scenes from screenplays or stage plays if you replaced the word “novelist” for “screenwriter” or “playwright”.
Copy the full text below, paste it into your AI of choice and replace [insert scene] with your draft.
[ROLE]
You are a Novelist-in-Residence with deep knowledge of narrative craft.
You prioritise psychological realism, scene economy, and subtext. No praise, no filler. Evidence-first.
[INPUT]
The user will supply a scene from their work-in-progress.
[SCOPE]
Task: Analyse the provided scene only. Do not rewrite or generate new prose.
[PROCEDURE]
1) Summarise the scene’s beats in ≤ 100 words.
2) Evaluate the scene against the Rubric.
3) Provide margin-style notes with direct text quotes.
4) Suggest one targeted revision strategy (not new prose).
[RUBRIC]
A) Clear external goal in-scene.
B) At least one irreversible turn.
C) Concrete sensory detail tied to action.
D) Subtext felt but unspoken.
E) No cliché diction.
[OUTPUT CONTRACT]
Return sections in this order with markers:
==BEAT-SUMMARY==
==RUBRIC-EVALUATION==
==MARGIN-NOTES==
==REVISION-STRATEGY==
[insert scene]
What the Engine Sees / What it Misses
What it sees:
The engine is ruthlessly clear on mechanics. With an example text, it flagged the scene’s strength in concrete detail and subtext, but also exposed its dramatic inertia: “the protagonist observes and reflects, but takes no external action”. The “weak” goal and “absent” turn verdict was unsentimental and correct. The revision strategy was equally sharp – add a simple beat of external pressure to break stasis.
What it misses:
The “misreading” wasn’t a flaw in logic but a limit of the rubric. The model reads for turns. It can’t see that in psychological realism, paralysis can itself be the event. A rubric designed for plot progression will flag such a scene as “weak,” even when that stasis is thematically necessary. The engine’s verdict wasn’t a mistake – it was an accurate reflection of the brief.
What does this mean?
The point of the analysis engine isn’t to hand down a verdict, it’s to hold your rubric up against your work and send an analysis back to you. The prompted model becomes a moderator of an editing standard. Its job is to keep you honest to your own brief. It diagnoses with precision, then hands the report back to the writer. But only you know the intent of your work. If your scene’s stasis is the heart of what you’re trying to show then you must stick with your vision. That final layer of judgement is where your art lies.
The true power of the Analysis Engine prompt is its modularity. By swapping a single component, you can ask an entirely different set of questions of the same text. After analysing the same scene using the Character Engine (see below), the feedback ignored plot mechanics and instead mapped the protagonist’s hidden desires against her misbeliefs and private fears. The diagnosis shifted from what was happening to why it was happening. That lens was far more useful to its author.
An LLM prompted into an analytical method cannot tell you if your work is “good.” It can only tell you if your work meets the criteria you gave it. The engine’s purpose is to show you the cracks in your craft - but only you can know if those cracks are a flaw in the foundation or a deliberate, essential part of the design.
working in dialogue: a three-step loop
Prompting is rarely a one-shot exercise. The most valuable experiences often come from iterating on cycles of testing, challenging, and refocusing. First, you run the Analysis Engine on your scene with the rubric you’ve chosen. Once you have its claims and evidence, you reflect and respond. This starts a dialogue to help you objectively understand your work as it stands.
If you fundamentally don’t agree with the output, challenge it. Ask the model to defend the opposite reading, to give a counter-argument with direct quotes, and even to rate its own confidence. It will be able to show you alternative ways the same passage might be read. Deep insights can come from forcing contradictions.
If the output feels like it doesn’t fit your needs, adjust the rubric. Perhaps move from scene mechanics to character tension (see ‘for power users’ below). Ask for a one-page revision plan instead of another verdict. Pull the feedback closer to your real intention for the draft.
Cycle through these stages until the diagnosis and your intent are aligned. At that point, the model has done its job: not necessarily by giving you an answer, but by sharpening the questions you’re asking about your own work. At its best, it will help expose something that hadn’t occurred to you and at its worst it will give you the language you need to talk about your work.
practical notes
Keep inputs manageable. The engine works best on one scene or short chapter, up to about 2,000 words. Longer chunks blur detail and weaken the analysis.
Remember that some passages will “fail” the default rubric by design. Reflective lyric writing, deliberate stasis, or language-driven micro-forms may all be marked down for lacking a clear goal or turn. That doesn’t mean the passage is broken – it means you need to change the question. Swap in a rubric that fits the form.
Finally, treat the Analysis Engine as a framework, not a straitjacket. You can adapt it to your own priorities, genre, or voice, and build personalised prompts that reflect your style. I’ll cover more on how to iterate on your prompting in future posts.
for power users
For those who want to see how truly modular this system can be, the key is to think of the prompt as a socket. By swapping out a single module, you can tilt the engine toward a different question.
Examples:
Character Engine – map tensions
Swap [RUBRIC] for: A) desire vs need B) misbelief C) public mask D) private fear E) pressure sources F) cost at climax.Structure Mapper – audit beats
Swap [PROCEDURE] for: 1) step through beats 2) check for visible choice 3) escalation 4) consequence 5) genuine turnLine Edit Pass – sentence precision
Swap [OUTPUT CONTRACT] for: paired “original > tightened” lines with reasoning.Risk Pass – challenge the safe draft
Swap [OUTPUT CONTRACT] for: list three editorial “pressure points” framed as options, not rewrites.
Think of the core engine as a framework. Each module slots into it as an interchangeable unit. The structure stays the same – only the questions it asks change.
closing thoughts
LLMs don’t know what makes a scene work. They can estimate an average from patterns and qualitative statements in their training data, or they can test against the criteria that you give them. A vague prompt will give you vague reassurance. A precise brief, built on your own intent, will give you a clear analysis.
The power lies in disciplined prompting. A well prompted chat acts as a professional service you would pay for. Done well, the model can give you “better than human” feedback and also save you money.
reader prompt
Run the engine on a scene from your draft. Where did it sharpen your intent, and where did it misread you? Share your results - I’d like to see how the model reflects your own practice.