5.2 Prompting the Persona: Template, Not Improvisation

Improvised prompting works. It is also the reason two posts from the same persona look like they came from different accounts.

The alternative is a template with fixed slots and frozen technical specification. It costs an hour to build and saves the same hour every week, while removing the largest source of unplanned variation in your output.

The prompt has six parts

In order, because the order carries information about what matters:

1. Identity. Fixed. The trigger token from 4.5, plus the reference conditioning. This section never changes except when the identity version changes, and then it changes everywhere at once.

2. Scene. The variable part: location, wardrobe, action, props, time of day. This is what you actually write per post, and it is the only genuinely creative slot.

3. Light. From the kit's three conditions (2.5). One per prompt, never mixed.

4. Lens and framing. Focal look, shot size, camera height — all fixed vocabulary from the kit.

5. Style and grade. The look, matching the saved preset from 2.5. If your grade is applied in post-processing, this section says only the photographic register, not the color.

6. Quality and texture. The retention phrases below.

The retention phrases, and what they actually do

Three phrases appear in almost every working prompt for this purpose:

• "preserve facial identity"
• "realistic skin texture"
• "natural skin pores"

What they do. They bias the model away from the two failure modes that cost the most: identity drift and over-smoothing. Pore-level texture is the single most effective guard against the plastic look from 1.4, because the model cannot render texture and a smooth surface at once — asking for texture forces detail into the face.

What they do not do. They do not make the face match a specific person. That is the reference layer's job, and no phrase substitutes for it. An operator who removes the reference and keeps the phrases will get a photorealistic stranger.

[UNVERIFIED] The specific effect of any single phrase varies across models and versions, and the effect sizes are reported by practitioners rather than measured. Treat them as cheap insurance with a plausible mechanism, not as guaranteed levers.

Freeze the technical specification

The distinction that reorganizes your prompt file:

Creative variables — scene, wardrobe, action, prop, time of day. These change every post, and they are what you should be thinking about.

Frozen specification — lens, framing rules, light condition vocabulary, quality phrases, negative prompt, output resolution. These do not change between posts. They are configuration.

Write the frozen part once, put it in a file, and build prompts by filling a slot. Two consequences: you stop re-deciding the same technical questions hundreds of times, and when you learn something about a technical setting, you change it in one place and it applies to every future post.

Negative prompts: what helps

A starting set, with reasons:

Negative prompts: what harms

Long lists of contradictory style words. "Photorealistic, cinematic, anime, oil painting" is not a filter; it is a conflicting instruction that flattens the result toward the average of all four.

Negatives that fight the scene. "Indoor" in the negative while the prompt places her in a kitchen produces a room with no room in it.

Negatives naming body parts in a positive context. A negative on "elbows" in a portrait prompt sometimes costs you the elbow.

Copying another operator's negative list. They are calibrating a different base model, a different adapter, and a different dataset. Negatives are local.

The practical approach: start with the table above, and add a negative only when you can point at a specific defect it would have prevented. A negative list that grows without evidence is a list that is fighting itself.

Templating: the 40-variant file

This is the operational core of the chapter.

Keep one file with forty scene variants — location, wardrobe, action, props, time of day — written out in full, each one a complete prompt except for the frozen specification. Producing a post becomes: pick a variant, paste it into the frozen frame, generate.

Why forty, and why a file:

Forty variants cover roughly two weeks at three posts a day, which means you never write a prompt under time pressure. Prompt quality collapses when you are publishing today.

A file makes the prop closet real. Chapter 2.3's twenty places with three objects each becomes forty concrete scenes. This is where that section pays off, and if you skipped it, this is where you notice.

It separates generation from invention. Writing forty prompts is a creative session; picking one is an operational step. Doing both at once produces neither well.

It makes batching possible. Chapter 5.4's batch pipeline consumes variants in groups. A prompt list is a batch manifest.

Refill the file quarterly, and refill from the review slot in 2.6 — where you have data about what performed rather than what you feel like making.

Per-platform prompting

The same persona needs different prompts on different surfaces, and the differences are structural rather than cosmetic. Illustration only, and platform specifications change constantly — verify current requirements before relying on them. [AS OF 2026-10]

Instagram feed. Square or 4:5, closer framing, the face readable at thumbnail size, warm grade, more emphasis on skin texture and styling detail. Feed stills are where attachment forms (2.6), so composition and wardrobe matter more than novelty.

TikTok and Reels covers. Vertical, and the frame will be seen at speed, small, often with a text overlay covering the lower third. Composition must survive a thumb: a single subject, high contrast, no fine detail in the lower third, and room in the frame for the caption.

Ads and promotional assets. Vertical or square, product visible, the persona's face in the upper third, negative space for text. Promotional imagery is judged by click behavior, not by beauty, and the grade is often brighter and flatter than your feed look — a deliberate deviation from the kit that should be recorded as such.

Stories. Vertical, looser composition, more informal styling. The frame will be seen for seconds and often annotated over, so precision matters less and immediacy matters more.

The recurrence across all four: framing and negative space change, identity does not. If you find yourself adjusting the identity section per platform, you have made a mistake one layer up.

A worked prompt, annotated

[identity]  <trigger token>, <adapter/ref conditioning>, woman, mid-20s
[scene]     standing in a small kitchen, morning, holding a mug,
            wearing an oversized grey knit sweater, marble counter,
            a plant and a laptop on the counter
[light]     soft window daylight from the left, overcast, no harsh shadow
[frame]     50mm-equivalent look, medium shot, eye-level camera
[style]     documentary photography, available light, unretouched
[quality]   preserve facial identity, realistic skin texture, natural skin pores
[negative]  blurry, low quality, extra fingers, malformed hands, asymmetric eyes,
            plastic skin, doll-like, watermark, text, overexposed

Six things to notice.

The scene section is where all the information is. Fourteen of the roughly thirty words describe the world. That ratio is correct: the identity and technical sections are boilerplate, and the scene is the work.

No facial description anywhere. Not one adjective about the face. This is the rule from 5.1, and it is the most important line in this chapter.

One light. "Soft window daylight from the left" plus "overcast" clarifies direction without introducing a second source. Two light sources in a prompt produce the impossible-shadow tell from 1.4.

Props carry the scene. A mug, a plant, a laptop. Props are how a generated kitchen becomes a specific kitchen, and they come from the prop closet.

The style section is short. "Documentary photography, available light, unretouched" does more than a sentence of adjectives, and it points at the same register the kit's preset will finish.

The grade is absent. It is applied in post (5.5) so it can be applied identically to every frame, including the ones you fix.

Seeds and reproducibility

One technical habit that separates a hobby from a pipeline: record the seed with the prompt.

Generation is stochastic. The same prompt, model, and reference set produce different frames on every run. A seed fixes that sample. With the seed recorded, a frame you liked is reproducible — which matters the moment you need a slightly different crop, or a colleague asks how you made something.

Three practical notes.

Hosted consumer tools often do not expose the seed. If you are on that tier, accept that your outputs are not reproducible and treat each good frame as a unique asset to archive carefully.

Node-based and scripted routes do expose it, and the seed belongs in the same manifest as the prompt variant (5.4). A manifest with prompts but no seeds is a description of a batch you can never run again.

Do not treat a seed as a guarantee. A model update, a driver change, or a library version bump can change output from the same seed. Reproducibility is within a configuration, not across configurations — the same reason adapters belong to their base model (4.5).

Where this goes wrong

Improvised prompts per post. The largest single source of inconsistency, and the easiest to fix.

Facial description in the prompt. Two identity sources competing.

A negative list that keeps growing. It will start fighting the scene.

Two light sources. The impossible-shadow tell.

No prompt file. You are writing forty prompts under deadline, badly.

A style section doing the grade's job. Post-processing exists so the grade is identical everywhere.

Adjusting identity per platform. The identity layer is platform-independent; only framing is not.

The gate

1. Does your prompt have six parts, in order, with identity fixed?
2. Are the retention phrases present in every prompt — and is the reference layer also present?
3. Is there a frozen specification file you are not re-deciding per post?
4. Does your negative list have a reason per entry, and no contradictions?
5. Do you have a forty-variant scene file, and is it drawn from the prop closet?
6. Do you have per-platform framing variants with an unchanged identity layer?

What you leave with

A six-part prompt structure with identity in a fixed slot, retention phrases with an honest account of what they do and do not do, the creative-versus-frozen distinction that turns prompting into configuration, negative-prompt tables for what helps and what harms, a forty-variant scene file that makes the prop closet operational, per-platform framing guidance with a stable identity layer, and one worked prompt annotated line by line.