| name | Grace Liu: Direction Discipline (Taste Doctrine) |
| description | The judgment layer above all craft — Human/Translation/Machine layer diagnosis, the authored-vs-mass-produced test, and the tier ladder for creative work. A RUBRIC other skills consume before they generate, not a workflow. |
| version | 1.0 |
| format | completion-engine |
| type | rubric |
| workflows | 0 |
Grace Liu: Direction Discipline
Creative Director, filmmaker, founder of Grace Liu Studio (NY/LA) and Human Directed™.
13 years directing image and video campaigns at Calvin Klein, David Yurman, Laura Mercier;
also Burberry Beauty, Mazarine, Chandelier Creative. Grew Calvin Klein's Instagram from
500K to 15M+ through CLIO- and Webby-winning campaigns including #mycalvins.
7M+ organic views on her own AI work since turning her practice toward it.
What she adds that nobody else in this house does: every other creative master here
operates inside a creative act — St. Pierre on art direction, Flynn on image operations,
Clark on cinematic craft. Liu is the only one who supplies the test for whether there was a
direction at all. She is the layer above the craft layer.
Read this first
This is a rubric skill, not a workflow skill. It has no workflows, and that is deliberate.
Liu's published archive is thin (two Substack essays, one Forbes panel piece, studio and program
site copy — roughly 3,500 words of her own language). She has published a doctrine, not a
procedure. Her procedure is taught live and privately, and none of it is public.
So this skill ships what the sources actually carry: a judgment rubric, quoted verbatim and dated,
that other skills load before they generate. Any step-by-step you find below is composed by this
house from her stated principles and is labelled as such. Nothing here reconstructs her program.
The one-line thesis
"The eye is already there. Prompting is a tactic. Direction is a discipline."
— Forbes, 2026-05-26
The core claim
AI slop is a direction problem, not a model problem.
Forbes, 2026-05-26: "Grace Liu believes most 'AI slop' is fundamentally a direction problem."
The corollary is what makes it operational. If slop is a direction problem, then no amount of
model-switching, prompt-tuning, or reroll volume fixes it — and every hour spent there is spent
at the wrong layer. The diagnosis has to name a layer before it names a fix.
Her second claim explains why this is newly urgent rather than merely true:
"AI collapses the distance between vision and execution. And in doing so, it exposes any
existing gaps in taste, medium fluency, or thinking… AI removes the insulation layer that once
separated taste from output. Meaning: you can no longer hide behind vendors, crews, budgets, or
production value. The output becomes a direct reflection of your mind."
— Substack, 2026-04-15
Production value used to launder weak direction. It no longer can.
THE RUBRIC — Part 1: The Three Layers
Her framework, as stated to Forbes (2026-05-26), separates AI production into three layers:
| Layer | Her definition (verbatim) | What lives here | The failure signature |
|---|
| 1. Human | "where intention and creative vision originate" | The idea. What is being reframed and for whom. What's in tension. The cultural position. Why this deserves to exist. The invisible planning decisions. | The work is competent and forgettable. Nothing is wrong with it. Nobody would miss it. It cannot be described except by describing its style. |
| 2. Translation | "where intent becomes machine-readable" | Articulation. Turning an eye into instructions precise enough that a crew — or a model — executes your vision. Reference discipline, named decisions, specification of everything you care about. | You got it right once and cannot recreate it. Or: the brief is adjectives, and the output is a plausible answer to a question you didn't actually ask. |
| 3. Machine | "where systems execute production" | Models, platforms, techniques, workflows, post. | The brief is specific, and the render disobeys a specific named element. Genuinely rare. |
And the constraint that makes it a hierarchy rather than a pipeline:
"The hierarchy never shifts. The human is always the director."
— Forbes, 2026-05-26
The Ladder Rule (house-composed from her doctrine — labelled, not hers)
You may not diagnose at a lower layer until the layer above it passes.
This is the whole anti-slop mechanism, because the reflex is always to start at Machine — it's the
only layer with buttons. Run it top-down and stop at the first failure:
- Human Layer gate. State, in one sentence each: what this is about, what is being reframed,
for whom, and what is in tension. If any cannot be stated, the diagnosis is complete. Do not
proceed. No craft fix, no model, no prompt reaches this.
- Translation Layer gate. Read the brief or prompt back. Could it describe a hundred different
acceptable outputs? Name every element you decided; everything unnamed was decided by the model.
If the specification is thinner than the intention, the failure is here.
- Machine Layer. Only now. And only for elements that were specified and disobeyed.
Stopping rule: a diagnosis that names failures at all three layers has diagnosed nothing. Name
the highest one and fix that.
THE RUBRIC — Part 2: The Five Tests
Five checkable questions, each anchored to a verbatim source. These are the actual gate.
1. The authored test
"Luxury creative has always been defined by invisible decisions during planning. The ones that
make an image feel authored rather than mass-produced." — Forbes, 2026-05-26
Ask: name three decisions made during planning that a viewer will never see but will feel.
If you can't name three, the work is mass-produced no matter how expensive it looks.
2. The defaults test
"…so the output consistently reflects your vision, not the model's defaults." — humandirected.ai, accessed 2026-08-02
Ask: what in this output did you decide, and what did the model decide for you?
Every element you didn't name is a default. Defaults are the actual texture of slop.
3. The insulation test
"You can no longer hide behind vendors, crews, budgets, or production value. The output becomes
a direct reflection of your mind." — Substack, 2026-04-15
Ask: strip the production value. Is there anything left? Does this reflect a specific mind, or
a general competence?
4. The multiplier test
"AI is a multiplier. It can multiply expertise, or it can multiply ignorance." — Substack, 2026-04-15
Ask: which is being multiplied here? Volume of output from a shallow position is not progress;
it is ignorance at scale.
5. The articulation test
"As creatives, we show. But AI requires us to articulate — with the precision of a director
leading a crew." — humandirected.ai, accessed 2026-08-02
Ask: could a competent stranger execute your vision from this brief alone? "Showing" a
reference and hoping is not direction. The precision bar is a director briefing a crew, not a
person describing a mood.
THE RUBRIC — Part 3: The Tier Ladder
From Substack, 2026-04-15 — "The Three Tiers of Creatives Using AI." Use it to grade the work,
not the person; the same practitioner produces at different tiers on different days.
| Tier | Her name for it (verbatim) | Requires | Her verdict |
|---|
| 1 | Prompt Operators | Tool knowledge | "Capable of generating practical assets like product images with precision. Useful, but increasingly replaceable as tool literacy becomes standard." |
| 2 | Aesthetic Stylists | Medium taste + tool knowledge | "Possess enough taste and technical control to make work feel elevated. Beautiful outputs, but often trend-sensitive and technique-driven." |
| 3 | World Architects | Worldbuilding, on top of medium taste + tool knowledge | "Build coherent symbolic universes rooted in classical medium principles, yet expressed through AI." |
Tier 3 is governed by three things she names explicitly:
- Consistent visual language
- Cultural and emotional intelligence
- Rich narrative and symbolic layers
"They know how to connect what is historically resonant with what feels urgent now. They
understand the emotional unlock that turns viewers into sharers." — 2026-04-15
Why the ladder is economic, not vanity:
"In an AI-saturated world, anyone can generate a passable luxury-looking image. Execution
becomes table stakes. The differentiator is worldbuilding — a coherent visual universe, anchored
in cultural context, that sustains meaning across time. Tier 3 becomes the world brands want to
enter." — 2026-04-15
Diagnostic use: most work that gets called slop is competent Tier 1 output being asked to do
Tier 3 work. Naming the tier gap is often the entire finding.
THE RUBRIC — Part 4: The Travel Test
For work meant to circulate — spec pieces, campaign concepts, anything whose job is to be shared.
From Substack, 2026-03-12, derived from her Gucci #GucciPrimavera piece (9 seconds, 100K views,
and — her actual quality signal — "notable fashion editors, underground cultural figures, fine
artists and general fashion lovers alike from different corners of the globe.")
Her four mechanics, verbatim-anchored:
-
The unlikely combination.
"The tension between the references is the idea. You're not just making a pretty
reinterpretation of a classic. You have to generate one that creates friction."
Check: name the two things that don't belong together. If everything in the piece agrees
with everything else, there is no idea — only a style.
-
Meets a conversation, doesn't chase a trend.
"There's a difference between work that chases a social trend, and work that meets a
cultural conversation. The former can feel unoriginal, while the latter feels inevitable."
Check: does this feel inevitable or reactive? Reactive is the tell.
-
The human crack in the marble.
"Elevated fashion creative typically stays cold. It aspires, but doesn't invite… The most
circulated work usually has one of these. A place where the marble cracks and something
universally human shows through."
Check: point at the one warm, funny, or human detail. If elevated work has none, it will
be admired and not shared. In her piece it was Mercury grinning and dropping his orange —
"an emotional puncture that converted viewers into sharers."
-
Layered meanings = multiple entry points.
"The work didn't simplify itself to reach a wider audience. It layered meanings so that
different readers could find their own entry point."
Check: name two different audiences and what each one gets. Note the anti-pattern she is
explicitly rejecting: do not simplify to widen reach. Layer to widen reach.
Her summary principle:
"Your strongest creative work isn't a technical flex or even a consistent creative signature —
it's a layered yet accessible cultural reinterpretation. It takes something that already lives
in people's minds and repositioning it somewhere they didn't expect. The resonance lives in that
repositioning. The medium used, AI or otherwise, is almost besides the point. The question is
always: what are you reframing, and for whom?" — 2026-03-12
"What are you reframing, and for whom?" is the single most portable question in this skill.
It is the Human Layer gate in six words.
How to run a direction-vs-prompting diagnosis (the short form)
On any creative task, output, or brief — five minutes, in order:
1. What are you reframing, and for whom? → Human Layer. Can't answer = stop here.
2. What's in tension? → Human Layer. Nothing = style, not idea.
3. Name 3 invisible planning decisions. → Human Layer. Can't = mass-produced.
4. Read the brief back — could it describe 100 → Translation Layer.
acceptable outputs? What did the model decide?
5. Was a NAMED element disobeyed? → Machine Layer. Only now is it a tool problem.
Then: name the highest failing layer, name the tier gap if there is one, and give one fix at
that layer. A fix proposed at a layer below the failure is the error this skill exists to prevent.
How other skills consume this
This rubric runs before generation and above craft. It does not replace any craft skill.
| Skill | Relationship |
|---|
skills/nick-st-pierre | Art direction = her Translation Layer, in depth. Liu says articulate; St. Pierre supplies the vocabulary and the sweep discipline to articulate with. Run Liu's Human gate, then St. Pierre. |
skills/rory-flynn | Image operations = Translation Layer systematized. Flynn's own warning ("systematize mediocrity") is a Human Layer failure by another name — a system built on no direction industrializes defaults. |
skills/dave-clark | Clark's cause #8, Stake ("Why does this exist? What is anyone supposed to feel?"), is where his craft taxonomy ends and hers begins. Clark diagnoses the seven mechanical causes of flat; Liu diagnoses whether there was a direction under them at all. Chain: Liu Human gate → Clark causes 1–7. |
skills/cinema-worldbuilder-pro, skills/generate, any generation lane | Liu supplies the pre-prompt gate. Her defaults test is the same claim as the craft-map rule that no generation fires with a freehand prompt — arrived at independently, which is corroboration. |
| Content, brand, positioning work | The travel test and "what are you reframing, and for whom?" transfer intact to non-visual work. She says so herself: "The medium used, AI or otherwise, is almost besides the point." |
Her working method, where the sources show it
Only one piece is documented end-to-end (the Gucci Primavera video, Substack 2026-03-12). What it
shows, stated as observation rather than as her taught method:
- She watched Demna's first Gucci show and read his inspiration statement — she works from the
designer's stated intent, not from the images alone.
- She reached for a classical analogue (Botticelli's Primavera, c. 1477–82) rather than a
contemporary reference — historically resonant against urgently current.
- She anchored on a word doing double duty: "Primavera" carried the painting, the season, and
Milan Fashion Week timing simultaneously.
- She restaged an archetype as a scene: "intellect surrendering to beauty and desire — but
restaged as a god getting distracted by women at the club."
- She built the emotional puncture as a small physical beat (a grin, a dropped orange) that did
not break the visual language.
- She used AI as a proving ground: "AI is the sketchbook, the pitch deck, and sometimes the first
audience test." Prototype, pressure-test culturally, then translate to other mediums once
resonance is proven.
- Her success metric was breadth of audience type, not view count.
What is NOT in this skill (honesty ledger)
- No procedure for the Translation Layer. She names it; she does not publish how she runs it.
Load
nick-st-pierre or rory-flynn for that layer's actual craft.
- No Machine Layer content whatsoever. No models, no techniques, no post workflow. Her program
covers it; the public record does not.
- No prompt formulas, no shot vocabulary, no lighting language. Do not attribute any to her.
- A framework discrepancy, unresolved. Forbes (2026-05-26) reports three layers:
Human → Translation → Machine. Her program site (accessed 2026-08-02) shows four modules:
Human Layer, Machine Layer, Post Production, Video & Film — with no Translation Layer named.
This skill uses the Forbes three-layer version because it is the doctrine she articulated in her
own quoted words. Do not present the program's module list as the framework. See
source-notes.md.
Execution Prompts
Quick Reference
- Genius context: genius.md — 22 verbatim-anchored judgment patterns, 7 signature
moves, her quality rubric, and her voice profile. Load before running either prompt.
- Source notes & fidelity ledger: source-notes.md — every source dated,
every claim graded, the full public archive inventory, and two namesake-collision warnings.
Execution Prompts (structure-pure v2)
2 deterministic practitioner prompts — each carries an Output Contract, Output Skeleton, and Quality Gate. When a deliverable matches one, Read it and honor its contract instead of improvising the output shape.
- Grace Liu — Direction Discipline Audit —
skills/grace-liu/references/prompts-v2/direction-discipline-audit.md
- Grace Liu — Human Layer Brief —
skills/grace-liu/references/prompts-v2/human-layer-brief.md
Quick Reference
- Genius context: genius.md — 22 verbatim-anchored judgment patterns, 7 signature
moves, her quality rubric, and her voice profile. Load before running either prompt.
- Source notes & fidelity ledger: source-notes.md — every source dated,
every claim graded, the full public archive inventory, and two namesake-collision warnings.
Execution Prompts (structure-pure v2)
2 deterministic practitioner prompts — each carries an Output Contract, Output Skeleton, and Quality Gate. When a deliverable matches one, Read it and honor its contract instead of improvising the output shape.
- Grace Liu — Direction Discipline Audit —
skills/grace-liu/references/prompts-v2/direction-discipline-audit.md
- Grace Liu — Human Layer Brief —
skills/grace-liu/references/prompts-v2/human-layer-brief.md
Quick Reference
- Genius context: genius.md — 22 verbatim-anchored judgment patterns, 7 signature
moves, her quality rubric, and her voice profile. Load before running either prompt.
- Source notes & fidelity ledger: source-notes.md — every source dated,
every claim graded, the full public archive inventory, and two namesake-collision warnings.
Execution Prompts (structure-pure v2)
2 deterministic practitioner prompts — each carries an Output Contract, Output Skeleton, and Quality Gate. When a deliverable matches one, Read it and honor its contract instead of improvising the output shape.
- Grace Liu — Direction Discipline Audit —
skills/grace-liu/references/prompts-v2/direction-discipline-audit.md
- Grace Liu — Human Layer Brief —
skills/grace-liu/references/prompts-v2/human-layer-brief.md
Quick Reference
- Genius context: genius.md — 22 verbatim-anchored judgment patterns, 7 signature
moves, her quality rubric, and her voice profile. Load before running either prompt.
- Source notes & fidelity ledger: source-notes.md — every source dated,
every claim graded, the full public archive inventory, and two namesake-collision warnings.
Execution Prompts (structure-pure v2)
2 deterministic practitioner prompts — each carries an Output Contract, Output Skeleton, and Quality Gate. When a deliverable matches one, Read it and honor its contract instead of improvising the output shape.
- Grace Liu — Direction Discipline Audit —
skills/grace-liu/references/prompts-v2/direction-discipline-audit.md
- Grace Liu — Human Layer Brief —
skills/grace-liu/references/prompts-v2/human-layer-brief.md
Quick Reference
- Genius context: genius.md — 22 verbatim-anchored judgment patterns, 7 signature
moves, her quality rubric, and her voice profile. Load before running either prompt.
- Source notes & fidelity ledger: source-notes.md — every source dated,
every claim graded, the full public archive inventory, and two namesake-collision warnings.
Execution Prompts (structure-pure v2)
2 deterministic practitioner prompts — each carries an Output Contract, Output Skeleton, and Quality Gate. When a deliverable matches one, Read it and honor its contract instead of improvising the output shape.
- Grace Liu — Direction Discipline Audit —
skills/grace-liu/references/prompts-v2/direction-discipline-audit.md
- Grace Liu — Human Layer Brief —
skills/grace-liu/references/prompts-v2/human-layer-brief.md
Quick Reference
- Genius context: genius.md — 22 verbatim-anchored judgment patterns, 7 signature
moves, her quality rubric, and her voice profile. Load before running either prompt.
- Source notes & fidelity ledger: source-notes.md — every source dated,
every claim graded, the full public archive inventory, and two namesake-collision warnings.
Execution Prompts (structure-pure v2)
2 deterministic practitioner prompts — each carries an Output Contract, Output Skeleton, and Quality Gate. When a deliverable matches one, Read it and honor its contract instead of improvising the output shape.
- Grace Liu — Direction Discipline Audit —
skills/grace-liu/references/prompts-v2/direction-discipline-audit.md
- Grace Liu — Human Layer Brief —
skills/grace-liu/references/prompts-v2/human-layer-brief.md
Quick Reference
- Genius context: genius.md — 22 verbatim-anchored judgment patterns, 7 signature
moves, her quality rubric, and her voice profile. Load before running either prompt.
- Source notes & fidelity ledger: source-notes.md — every source dated,
every claim graded, the full public archive inventory, and two namesake-collision warnings.
Execution Prompts (structure-pure v2)
2 deterministic practitioner prompts — each carries an Output Contract, Output Skeleton, and Quality Gate. When a deliverable matches one, Read it and honor its contract instead of improvising the output shape.
- Grace Liu — Direction Discipline Audit —
skills/grace-liu/references/prompts-v2/direction-discipline-audit.md
- Grace Liu — Human Layer Brief —
skills/grace-liu/references/prompts-v2/human-layer-brief.md
Quick Reference
- Genius context: genius.md — 22 verbatim-anchored judgment patterns, 7 signature
moves, her quality rubric, and her voice profile. Load before running either prompt.
- Source notes & fidelity ledger: source-notes.md — every source dated,
every claim graded, the full public archive inventory, and two namesake-collision warnings.
Execution Prompts (structure-pure v2)
2 deterministic practitioner prompts — each carries an Output Contract, Output Skeleton, and Quality Gate. When a deliverable matches one, Read it and honor its contract instead of improvising the output shape.
- Grace Liu — Direction Discipline Audit —
skills/grace-liu/references/prompts-v2/direction-discipline-audit.md
- Grace Liu — Human Layer Brief —
skills/grace-liu/references/prompts-v2/human-layer-brief.md
