Amini One-Shot

---
name: Amini One-Shot
description: >-
  Compile one paste-ready Amini brand prompt for a tool that cannot read your files.
  Use for "NotebookLM prompt", "Copilot prompt for PowerPoint", "Pitch prompt", "Gamma
  prompt", "Canva prompt", "Google Slides / Gemini prompt", "make an on-brand deck in
  <external tool>", "one prompt with our brand rules", or "single-shot prompt". Merges
  the Amini brand kit (palette, type, layout), the humanizer voice rules, and the
  illustration direction into one block sized to the target tool's instruction box.
  Ships a compiler with a brand linter.
license: MIT
metadata:
  author: Amini
  version: "1.0.0"
---

# Amini One-Shot

NotebookLM, PowerPoint Copilot, Pitch, Gamma, and Canva all have the same shape: one box,
one shot, no filesystem. They cannot read `brandkit/DESIGN.md`, cannot run the palette
auditor, and will not ask a follow-up question. Whatever the box does not say, the tool
invents, and what it invents is a purple gradient with the word "seamless" on it.

This skill compiles that box. It merges three sources into one block:

| Source | What it contributes |
| --- | --- |
| `brandkit/DESIGN.md` | Palette hex, type, hierarchy, layout, the hard don'ts |
| `humanizer-voice` | Sentence rhythm, banned vocabulary, no summary endings |
| `amini-illustrations` | Flat vector direction, one Sunglow accent, the Ami figure |

Inside Claude Code, use the three skills directly. They read files, run checkers, and
iterate. This one exists for the moment the work leaves the terminal.

## Read these as needed

- `references/targets.md` — what each tool profile assumes and where each tool fights back.
- `references/assembly.md` — how the blocks compress, what to cut first, editing the source blocks.
- `references/aftercare.md` — what to check when the output comes back, and how to bring it home.
- `assets/blocks/` — the three block files. Edit these, not the compiled prompt.
- `assets/targets.json` — the tool profiles: blocks, tier, character budget, header, footer.
- `examples/ministry-briefing/` — one brief compiled to three targets at three sizes.

## Workflow

### 1. Get the brief

Four things, and the first is the only one that is required:

- **task** — the subject, in a sentence
- **audience** — who is in the room
- **sources** — what the tool has been given, by name
- **length** — slide count, minutes, or word count

If the user has an article, deck spec, or research paper open, pull these from it rather
than asking. Ask only for what you genuinely cannot infer.

### 2. Compile

```bash
python3 scripts/build_prompt.py --list-targets

python3 scripts/build_prompt.py --target pitch \
  --task "Sovereign AI infrastructure: not a telco, not SaaS, not public cloud" \
  --audience "Ministry of Digital Transformation, permanent secretary and CISO" \
  --sources "Barbados pilot debrief, Benin ingestion report, SIDSense paper" \
  --slides 11

python3 scripts/build_prompt.py --brief brief.json --target gamma --mode light -o prompt.txt
```

Paths are relative to this skill's directory. Stdlib only, Python 3.9+.

The prompt goes to stdout and the lint report to stderr, so `> prompt.txt` gives you a
clean file with the report still on screen. `--strict` exits non-zero for CI.

Useful flags: `--tier XS|S|M|L` to override the profile's size, `--mode light`,
`--blocks ground,voice` to drop a block the tool has no use for, `--budget` when you know
the real limit of the box.

### 3. Read the lint report before you hand it over

It fails on unfilled slots, on an em dash in a prompt that bans them, on a brand token that
got trimmed away, and on going over the character budget. A budget finding is the common
one, and the fix is to drop a tier, not to delete lines by hand.

Budgets are conservative defaults, not published limits. NotebookLM's customisation box is
genuinely small; a deck tool's box usually is not. If you know the real number, pass it.

### 4. Hand it over, then say what it cannot do

Give the user the block and the three-line version of what it will not fix. A pasted prompt
carries rules, not fonts and not files:

- **Fonts do not travel.** No tool will install Helvetica Now Display. Expect a substitute
  and plan to reset type on export.
- **The logo is placed, never generated.** Drop the SVG in afterwards from
  `brandkit/assets/logo/`.
- **Nothing is audited.** No palette check runs inside Gamma. The output comes back to a
  session with the real skills for that.

Anything the tool produces that will be reused, bring home: run the copy through
`humanizer-voice`, the images through `amini-illustrations/scripts/palette_check.py`, and
rebuild a deck that matters with `amini-decks` rather than exporting theirs.

## What this skill will not do

Do not paste a compiled prompt into a Claude session and call it done. In a session with
files, `amini-decks` builds a real `.pptx` and `amini-illustrations` audits every image.
A one-shot prompt is what you use when that is not available, and it is strictly worse.

Do not invent brand rules to fill a block. Every line traces to `brandkit/DESIGN.md` or to
one of the two skills. New rule, new pull request, and it goes in the block file so every
prompt gets it.

## Output register

Give the compiled block, the character count against the budget, and the limits. Do not
explain the brand rules back to the user. The block carries them.