# Build your own AI design agent

**The CV Method · The Playbook · No. 4 · Under the Hood · Aug 2026**
*by Corina · @the_cv_method · hello.thecvmethod.com*

Stand up a small AI worker that turns your copy into **on-brand visuals** — the same slides every time, no drifting fonts or off-brand colors. Here's the pipeline I actually use to design my posts.

---

## What you're building

Not "ask AI to make me a pretty slide" — that drifts every single time, new fonts, new colors, a different vibe. Instead: **lock your brand into a fixed template, and let the agent only fill in the words.** The design is decided once; the agent pours approved copy into slots it can't move or restyle. A hundred posts come out looking like one brand — because they're the same template, every time. It rests on one rule.

## Rule zero — the brand is the source of truth. The agent fills the words; it doesn't redesign.

The agent never invents styling. It maps your approved copy into a locked template — same layout, palette, and fonts every run. That single constraint is what makes it consistent instead of a slot machine.

## Step 1 · Pick your stack

Four pieces. Here's what I use and the general equivalent — swap in whatever you already have.

| Piece | What it does | I use | Swap for |
|---|---|---|---|
| **The brand system** | Your colors, fonts, and spacing in one place the template reads. | A tokens file (CSS variables) | A Canva Brand Kit · Figma variables |
| **The template** | A fixed layout with named text slots (cover · content · stat · CTA). | An HTML template kit | Canva templates · Figma frames |
| **The renderer** | Turns a filled template into an image at exact size. | A headless-browser screenshot | Canva/Figma export |
| **The agent** | Maps your approved copy → the slot text, one entry per slide. | Claude + a small script | Any tool-use LLM |

## Step 2 · Lock the brand into tokens

Define your palette, fonts, and spacing as **tokens** in a single file the template reads — say a navy base, a green accent, one bold pop color, and your three fonts. Every template pulls from these, so nothing is decided ad-hoc.

Change a token, and every future post updates with it. This "decide it once" step is what actually kills the drift — the agent literally has no other colors to reach for.

*Tip: keep one gradient moment per slide, and never put body text on a live gradient — it reads as brand polish and stays legible.*

## Step 3 · Build fixed templates with named slots

Design a few slide types by hand — **cover, content, stat/quote, CTA** — and give each one **named text slots** (headline, body, footer). That's it. The agent's only job is to put the right words in the right slots.

Because it can't move elements, choose fonts, or recolor anything, it **can't go off-brand** — the worst it can do is write weak copy, which you catch on review. Design the fence once; pour words into it forever.

*Tip: four to six slide types cover almost everything. Add a new type deliberately, not once per post — that's how the system stays tight.*

## Step 4 · Feed it approved copy → a spec

Hand the agent your **approved** caption. It returns a simple **spec**: which template each slide uses, and the text for every slot — one entry per slide. Plain and small enough to read at a glance before anything renders.

Two honest rules ride here: any client proof uses a **category label**, never a real name; and a stat slide carries **real numbers only** — the agent never invents a figure to fill the space.

*Tip: keep the spec human-readable (a little JSON or a table). Eyeballing the spec is faster than re-rendering — you fix wording before you spend a render.*

## Step 5 · Render at exact size, review, ship

Render each filled slide to an image at the **real dimensions** (1080×1350 for a carousel), so what you see is exactly what the feed shows. Then **you review the images** — the human gate — and post the ones that land.

The agent drafts the visuals; you approve them. Same as every good AI worker: it proposes, you ship. Nothing posts itself.

*Tip: rendering is deterministic — the same spec makes the same slides — so what you approve is precisely what goes out. No surprise re-interpretation.*

## Keep it honest & on-brand

- The brand system is the source of truth.
- On-palette only · one gradient moment.
- Real numbers only — no invented stats.
- Category labels, never a client's name.
- Approved logos, clear-space respected.
- You approve every image before it posts.

## Steal these prompts

**Copy → slide spec — hand the agent your approved caption**

```
Here's my approved caption: [paste].
Turn it into a slide spec for my template. For each slide, output: template type (cover / content /
stat / cta) and the text for each named slot (headline, body, footer).
Rules: use ONLY my brand template — don't invent styling; 5–8 slides; category labels for any client
proof (no real names); real numbers only (never invent a figure to fill space).
Return it as a simple list I can eyeball before I render.
```

## The starter checklist

1. Put the brand in **tokens** — colors, fonts, spacing.
2. Build **4–6 fixed slide templates** with named slots.
3. One **gradient moment** per slide; no body text on a gradient.
4. Agent maps approved copy → a **slide spec** (template + slots).
5. **Category labels** for proof; **no invented numbers**.
6. Render at **exact size** (1080×1350 carousel).
7. **Review every image** — the human gate.
8. Post — the agent drafts visuals, **you approve**.

---

**Want a branded design agent — or the whole content system — built for you?** I set up the tokens, the templates, and the render pipeline so your posts look like one brand every time — then hand you the keys. Start at **hello.thecvmethod.com**.

*The CV Method · Build your own AI design agent · No. 4*