feat(claude-design): add 01-prompt-fundamentals five-layer stack reference

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
This commit is contained in:
Kjell Tore Guttormsen 2026-05-18 12:13:41 +02:00
commit ce715ef087

View file

@ -0,0 +1,265 @@
# Prompt fundamentals — the five-layer stack
**Last updated:** 2026-05-17 | **Verified:** research/03-prompt-patterns-intent-presets.md
**Status:** Beta (Labs research preview)
**Captured-on date:** 2026-05-16
This file documents the universal prompt framework an operator applies across every Claude Design intent preset. The five layers compose into one prompt block. Layers 1 to 3 are load-bearing for every preset; layers 4 and 5 are the refinement turn.
Every authoritative claim cites an Anthropic primary source. Where community practice extends an Anthropic concept, the extension is labelled and attributed.
---
## Layer 1 — Goal / Layout / Content / Audience (GLCA)
Anthropic's verbatim framework for every Claude Design prompt. The framework is published in the get-started support article `https://support.claude.com/en/articles/14604416-get-started-with-claude-design`. Anthropic's framing: a good Claude Design prompt names the **Goal**, the **Layout**, the **Content**, and the **Audience**, in that order, before any aesthetic specification.
The four canonical questions:
- **Goal** — what is the artifact for? "An admin dashboard for monitoring API latency", "an onboarding flow for first-time users", "a landing page that converts free trial signups". One sentence.
- **Layout** — what is the page structure? Header / hero / metrics row / table / footer; or: hero / three-feature-grid / pricing table / CTA. Name the regions.
- **Content** — what fills the regions? Real data placeholders if you have them, named labels if not. Avoid generic "lorem ipsum" — the model defaults to convergent middle-ground content if you do not constrain it.
- **Audience** — who reads or uses this artifact? Internal team, external stakeholder, B2B procurement, B2C consumer, investor. Audience determines tone, density, and aesthetic.
Anthropic publishes three verbatim canonical examples in the same support article:
```
Goal: An analytics dashboard for our customer success team
Layout: Top metrics row (4 KPIs), main chart panel, recent activity table
Content: Today's MRR, 30-day churn, NPS, expansion revenue; revenue chart;
the last 10 account events
Audience: Internal CS leads — they're in this thing every day, want density
and signal, not flashy
```
```
Goal: A mobile onboarding flow for a new fitness app
Layout: Welcome screen, goal-selection (3 cards), motion preference, sign-in
Content: Headlines, single CTA per screen, accessible touch targets
Audience: First-time users, gym beginners, ages 25-45
```
```
Goal: A SaaS landing page that converts free trial signups
Layout: Hero, three-feature grid, social proof, pricing table, FAQ, footer CTA
Content: Product name placeholder "ProductX", real headline benefit copy,
three feature blurbs (icon + headline + line)
Audience: B2B technical buyers evaluating dev tools
```
The GLCA framework is sufficient on its own for a first prompt at intent preset `designs`. For other presets, GLCA composes with the per-preset pattern in `references/presets/<preset>.md`.
Source: `https://support.claude.com/en/articles/14604416-get-started-with-claude-design`.
---
## Layer 1.5 — Start simple, layer in complexity
The same Anthropic get-started article publishes verbatim incremental-prompting advice: do not ship a 600-word first prompt. Ship a 120-word first prompt that names GLCA, see what Claude Design produces, then add complexity in turn two and turn three.
Anthropic frames this as the dominant failure mode for first-time Claude Design operators: over-specifying the first prompt produces an output that is dense but generic. The remedy is staged — let Claude Design make its default choices, then react to what it produces with targeted constraints.
This frames how the rest of the stack composes. In turn one, ship layers 1 + 2a (or 2b) + 3. In turn two, add layer 4. In turn three, add layer 5 emphasis-weighting.
Source: `https://support.claude.com/en/articles/14604416-get-started-with-claude-design`.
---
## Layer 2a — Concrete-alternative-spec house-style control
The first of two Anthropic-documented house-style controls. Verbatim guidance from `https://platform.claude.com/docs/en/docs/build-with-claude/prompt-engineering/claude-4-best-practices`: name a concrete aesthetic family with explicit visual primitives rather than gesturing at a style.
The Anthropic-published exemplar — the AEFRM (Anthropic Engineering Field Reference Material) example — is verbatim usable:
```
Aesthetic family: industrial-utilitarian, slate-monochrome
Color palette (CSS hex):
--color-bg: #E9ECEC
--color-surface: #C9D2D4
--color-muted: #8C9A9E
--color-fg: #44545B
--color-ink: #11171B
Typography: square angular sans-serif (Söhne, Inter Variable as fallback);
no rounded glyphs; weight 500 for body, 700 for headers
Corner radius: 4px throughout — no fully rounded buttons, no pill shapes
Motion: transition: all 160ms ease-out on hover; no springy easing
Density: dense (table rows 32px tall; padding 8px on cards)
Surface: flat — no shadows, no glassmorphism
```
The control works because Claude Design reads this as a concrete brief and constrains its aesthetic decision space accordingly. Without an explicit concrete-alternative-spec, the model defaults to its convergent middle-ground aesthetic (rounded corners, generous spacing, friendly typography, gentle shadows — Anthropic's documented "AI-slop" default).
The hex palette, corner radius, and motion timing are all required — naming "industrial-utilitarian" alone is gesturing, not specifying.
Source: `https://platform.claude.com/docs/en/docs/build-with-claude/prompt-engineering/claude-4-best-practices`.
---
## Layer 2b — Propose-options-before-building
The second Anthropic-documented house-style control, also from `https://platform.claude.com/docs/en/docs/build-with-claude/prompt-engineering/claude-4-best-practices`. When the operator does not know exactly which aesthetic to brief, Anthropic publishes a verbatim prompt template asking Claude Design to propose four distinct visual directions before committing to one.
The verbatim prompt:
```
Before building the dashboard, propose 4 distinct visual directions.
For each, give:
- bg hex
- accent hex
- typeface (named, not gestured)
- one-line rationale tying the direction to the audience and goal
Wait for me to pick a direction before generating the artifact.
```
This forks the conversation: turn one returns four named directions, the operator picks one, turn two generates against the chosen direction. The cost is one extra round; the upside is the operator avoids the dead-end of generating against an aesthetic that does not fit and only finding out after generation.
Use layer 2b when layer 2a is not feasible (the operator does not yet know the aesthetic). Use layer 2a when the aesthetic is known.
---
## Layer 3 — AI-slop avoid-list (negative constraints)
Anthropic publishes a verbatim banned-items list in `https://claude.com/blog/improving-frontend-design-through-skills` and reinforces it in the open-source frontend-design skill at `https://github.com/anthropics/skills/skills/frontend-design/SKILL.md`. The list names the convergent middle-ground patterns that Claude Design defaults to when underspecified.
Anthropic's verbatim AI-slop fingerprints to avoid:
- **Typography slop:** Inter, Roboto, Arial as default body font. Space Grotesk is flagged as overused. Default to a concrete-named typeface in the brief, not a generic sans-serif.
- **Color slop:** purple gradients on white backgrounds; solid-color hero backgrounds; convergent middle-ground palettes (the muted blue-and-grey "professional" default).
- **Layout slop:** cookie-cutter three-column feature grids; centered-hero-with-CTA defaults; full-width-image-with-text-overlay defaults.
- **Motion slop:** scattered micro-interactions; bouncy spring easing on hover; pulse animations on idle elements.
- **Complexity-to-vision mismatch:** ornate components on simple layouts; flat components on otherwise rich layouts.
Operator-actionable copy-paste anti-prompt block (composes with layers 1 and 2):
```
Negative constraints — do not produce any of:
- Inter, Roboto, Arial, or Space Grotesk as the primary typeface
- Purple gradients on white backgrounds
- Solid-color hero backgrounds
- Three-column feature grids with icon + headline + line
- Centered-hero-with-single-CTA layout default
- Bouncy spring easing on hover transitions
- Pulse / breathing animations on idle elements
- Glassmorphism, neumorphism, or generic "modern SaaS" defaults
If you find yourself defaulting to any of these, stop and ask me to
clarify the aesthetic before continuing.
```
Sources: `https://claude.com/blog/improving-frontend-design-through-skills` and `https://github.com/anthropics/skills/skills/frontend-design/SKILL.md`.
---
## Layer 4 — Four design dimensions to optimize
Anthropic's verbatim per-dimension guidance from `https://github.com/anthropics/skills/skills/frontend-design/SKILL.md`. Four dimensions to brief explicitly when refining beyond the first turn:
- **Typography** — name typeface, modular scale (e.g., 1.250 minor third or 1.333 perfect fourth), weight palette, line-height palette, letter-spacing for headings. Anthropic's frontend-design SKILL.md publishes specific modular scales and weight palettes verbatim.
- **Color** — beyond palette hex, specify semantic roles (background, surface, accent, muted, error, success). Define interaction states explicitly (hover, active, disabled, focus). Anthropic's guidance: avoid relying on opacity for state changes; use explicit color tokens.
- **Motion** — name easing curves (ease-out, cubic-bezier values), name durations (120ms / 160ms / 240ms tiers), name what gets animated and what does not. Anthropic's guidance: motion should clarify hierarchy and confirm interaction; avoid decorative motion.
- **Backgrounds** — flat surface vs depth, when to layer surfaces, when shadows or borders define edges. Anthropic's guidance: backgrounds carry meaning; the bare-default-white background is rarely the right choice.
In turn two of an iteration, add layer-4 dimension specs to the brief. In turn three, refine the dimension that drifted most from intent.
Source: `https://github.com/anthropics/skills/skills/frontend-design/SKILL.md`.
---
## Layer 5 — Four design grading criteria
Anthropic publishes verbatim grading criteria for design quality in `https://anthropic.com/engineering/harness-design-long-running-apps`. Four criteria, used as emphasis weights:
- **Design quality** — does the artifact look intentional, not defaulted? Is the aesthetic coherent across regions?
- **Originality** — does the artifact avoid the convergent middle-ground? Does it surprise without being weird?
- **Craft** — does the artifact feel detailed and considered at every level — typography, spacing, alignment, hierarchy, color?
- **Functionality** — does the artifact work for its goal and audience? Would it survive a usability test or a stakeholder review?
Anthropic's emphasis-weighting recommendation: in the prompt, weight which criterion matters most for *this* artifact. A dashboard for internal use weights functionality and craft highest. A pitch deck for an external investor weights design quality and originality highest. A wireframe for early exploration weights functionality highest with craft and originality deprioritized.
Operator-actionable layer-5 block:
```
Grading criteria for this artifact, in priority order:
1. {craft|design quality|originality|functionality} — weight 0.4
2. {one of the others} — weight 0.3
3. {one of the others} — weight 0.2
4. {the remaining one} — weight 0.1
Optimize against this ordering. If the artifact has to trade off,
trade off the lowest-weighted criterion first.
```
The non-monotonic-improvement caveat applies — Anthropic notes that quality across iterations is not strictly increasing. If turn three is worse than turn two on a critical criterion, the recovery move is documented in `references/03-iteration-and-session.md` ("pivot to an entirely different aesthetic if the approach wasn't working").
Source: `https://anthropic.com/engineering/harness-design-long-running-apps`.
---
## How the layers compose into one prompt
A worked example for the `designs` intent preset, dashboard, three turns:
### Turn 1 — layers 1 + 2a + 3 (the first prompt)
```
Goal: An admin dashboard for monitoring API latency by route, by region,
and by P50/P95/P99
Layout: Header with environment switcher; top metrics row (4 KPIs:
global P95, error rate, throughput, active requests); main chart
(time series, P50/P95/P99 lines); routes table with sortable
latency columns; alerts sidebar
Content: KPI placeholders are real metric names; chart uses synthetic
24-hour data; table has 12 routes with realistic paths
(/api/v1/users, /api/v1/orders, etc.)
Audience: Platform engineers, on-call rotation, ages 25-45,
comfortable with dense interfaces
Aesthetic family: industrial-utilitarian, slate-monochrome
Color palette (CSS hex):
--color-bg: #E9ECEC
--color-surface: #C9D2D4
--color-muted: #8C9A9E
--color-fg: #44545B
--color-ink: #11171B
Typography: square angular sans-serif (Söhne, Inter Variable fallback);
no rounded glyphs
Corner radius: 4px throughout
Motion: transition: all 160ms ease-out
Density: dense (32px table rows, 8px card padding)
Surface: flat — no shadows
Negative constraints — do not produce any of:
- Inter, Roboto, Arial, or Space Grotesk as primary typeface
- Purple gradients on white backgrounds
- Pulse animations on idle elements
- Glassmorphism, neumorphism, generic "modern SaaS" defaults
```
### Turn 2 — add layer 4 dimensions
Operator reacts to turn-1 output by adding typography modular scale, semantic color roles, motion easing, and surface-depth rules.
### Turn 3 — add layer 5 weighting
Operator specifies that craft and functionality are the two highest-weighted criteria for this dashboard; design quality is third; originality lowest.
### Turn 4+ — Tweak panel takes over
Most subsequent refinements happen in the Tweak panel (per-artifact Claude-generated controls; zero-token surgical edits) — see `references/03-iteration-and-session.md` for the surface cascade.
---
## Source map
The five layers anchor on four Anthropic primary sources plus one open-source skill:
| Layer | Anthropic source |
|-------|------------------|
| 1, 1.5 | `https://support.claude.com/en/articles/14604416-get-started-with-claude-design` |
| 2a, 2b | `https://platform.claude.com/docs/en/docs/build-with-claude/prompt-engineering/claude-4-best-practices` |
| 3 | `https://claude.com/blog/improving-frontend-design-through-skills` + `https://github.com/anthropics/skills/skills/frontend-design/SKILL.md` |
| 4 | `https://github.com/anthropics/skills/skills/frontend-design/SKILL.md` |
| 5 | `https://anthropic.com/engineering/harness-design-long-running-apps` |
Re-research trigger: any of the four URLs returning 404 or shifting content materially; Anthropic publishing a sixth layer or revising any of the five. Captured-on date 2026-05-16 — the layer-1 framework has been stable since the launch announcement.