Direct

stardust:direct

Resolve the user's freeform redesign intent into a complete target specification: project-root PRODUCT.md and DESIGN.md (impeccable format), a DESIGN.json sidecar with the divergence audit trail, and a stardust/direction.md with the full reasoning trace.

direct produces the spec against which prototype and migrate operate. It never writes prototypes or migrates pages — those are downstream sub-commands.

Inputs

• <phrase> — optional positional. The user's freeform intent ("make it better", "more Linear less Salesforce", "feel more premium on a small screen"). If omitted, ask the user for one.
• --re-direct — optional. Replace the current direction with a new one. Triggers stale-flagging on prototyped / approved / migrated pages per skills/stardust/reference/state-machine.md. Default behaviour without the flag is additive: if a direction already exists, the agent asks before replacing.
• --rebrand — optional. Force rebrand mode (full divergence-seed roll, no Mode A inheritance) regardless of whether the captured brand surface has signal. The default mode is brand-faithful whenever _brand-extraction.json is signal-strong (see § Setup step 5 and Phase 2 § Mode-detection precedence); pass --rebrand to opt out explicitly when the user's phrase is ambiguous about whether they want a refresh or a clean-slate redesign.
• --prep — optional. Run in migrate-prep mode: confirm the type catalog, finalize the module catalog, capture color reservations and brand-level metadata defaults, re-evaluate direction against the wider crawl. See § Prep mode below. Typically invoked via the prepare-migration orchestrator.
• --add-variant <name> — optional. Add a new variant against the existing direction without re-running intent reasoning or mode detection. Writes DESIGN-<name>.{md,json} at the project root and appends a per-variant section to stardust/direction.md. The previous direction stays authoritative; existing prototypes are not stale-flagged. Used when variants are spec'd incrementally — typical workflow: render variant A, review, then ask for B and C as "the alternatives we discussed earlier." See § Add-variant mode below for the per-field inheritance rules.

Setup

1. Run the master skill's setup (skills/stardust/SKILL.md § Setup) — hard impeccable dep check, context loader, state read.
2. Verify stardust/state.json exists and contains at least one extracted page. If not, stop and recommend $stardust extract <url> first.
3. Read stardust/current/_brand-extraction.json. If absent, stop — extract did not complete brand-surface extraction; re-run extract. 3b. Cross-site brand inputs (when present). Two extract-written signals widen the brand surface beyond the primary origin:
   • state.json.designSource — design-donor mode (extract ran with --design-source <url>). The donor's derived system at stardust/canon-source/DESIGN.{md,json} becomes the target the Mode A pins bind to: palette and type pin to the donor surface while content, IA priorities, and per-page evidence stay with the primary origin. Surface in the plan: "Design-donor mode — target system inherited from ."
   • _brand-extraction.json.origins[] — sibling-property evidence captured via --brand-source. Traits captured on a sibling origin count as captured evidence for trait amplification (variant B/C briefs, improvements items) exactly like primary-origin evidence; cite the origin in the evidence citation. Under Mode A the pins still come from the primary origin unless designSource says otherwise — sibling evidence widens what can be amplified, not what is pinned.
4. Read stardust/direction.md if present. If a prior direction exists and --re-direct was not passed, ask whether the user wants to refine the existing direction or replace it.
5. Validate provenance on every in-scope page (prep mode only). When --prep is active, call validateProvenance(page) per skills/stardust/reference/state-machine.md § Provenance validation on every extracted page in state.json before typing or module detection. Abort with the helper's error when any page lacks live-render evidence. Surface Provenance OK on N pages in the prep summary. Discovery- mode direct does not validate (it operates on a 5-page sample for which extract's own write-time refusal is sufficient).
6. Classify the captured brand signal. Read _brand-extraction.json and stamp one of:
   • signal-strong — palette has ≥ 3 distinct colors (after near-duplicate clustering and excluding pure black/white if they are the only entries) AND at least one captured type family is named in type.headingFamily.name or type.bodyFamily.name. This is the common case for any extracted commercial site.
   • signal-thin — palette has 2 colors OR no captured type family OR type.scaleAudit.kind === "ad-hoc" with fewer than 3 distinct heading sizes. The brand exists but cannot fully anchor a refresh.
   • signal-absent — palette has 1 color or 0, or _brand-extraction.json._provenance.notes flags the extraction as failed / login-walled / iframe-dominated. The classification feeds the Mode-detection precedence in Phase 2. Surface the classification in the plan when it would change the default mode.

Procedure

Phase 1 — Reasoning

Run the full intent-reasoning procedure from skills/stardust/reference/intent-reasoning.md. Steps 1-6: restate the phrase in dimensional vocabulary, identify movement, identify gaps, ask at most two clarifying questions, map to an impeccable command sequence, show the plan to the user.

Worked examples in skills/stardust/reference/intent-examples.md calibrate the style. Hard ceiling on questions: two per turn, no exceptions.

Hands-off mode (per skills/stardust/SKILL.md § Hands-off mode, state.json.handsOff: true): ask nothing and wait for nothing. Derive every answer the questions would have collected from the captured evidence — density and ia-fidelity from their documented defaults and trigger conditions, audience and register from the captured surface — and record each as a named assumption in direction.md § Movements (e.g. density: balanced (hands-off default — multi-audience floor fired)). The plan is still written; execution proceeds without the confirmation gate. Question budgets and gates below that say "ask the user" resolve the same way: derive, stamp the assumption, proceed.

Density tuning (one-shot, only when unmoved)

When the user's phrase does not move the density axis (per reference/intent-dimensions.md § 4), and the resolved register is brand, ask one short follow-up — count it within the two-question ceiling:

Density tuning — (a) airy (NYT-Opinion-tier breathing, ~96px section padding), (b) balanced (calm but compact, ~64–72px), (c) packed (data-dense, ~40–48px). Default for brand-register sites with multi-audience IA is (b) balanced; pick (a) only when the page is editorial-led with deep per-section density.

If the user answers, stamp the chosen tier in direction.md § Movements as density: <tier>. If unanswered, default to balanced (not airy) for brand register and stamp density: balanced (default).

Skip this question entirely when:

• The user's phrase already moved density (any of "make it denser", "more breathing room", "compact", "tight", "spacious" count as movement).
• The register is product (default packed per § 4).
• The register is ambiguous and resolving it earlier in the reasoning is the higher-value question — defer density to the next turn rather than burning a question slot.

The tier propagates to DESIGN.md's spacing.sectionPadding deterministically per intent-dimensions.md § 4: airy = 96px, balanced = 64px, packed = 48px. Phase 4 picks the value from this stamp without re-asking.

IA-fidelity tuning (one-shot, only when unmoved)

When the user's phrase does not auto-pin ia-fidelity (per reference/intent-dimensions.md § 9 — i.e. none of "verbatim", "same IA", "keep the structure", "swap the surface", "don't rethink the IA" and none of "reimagine", "rethink", "deeper redesign", "what if" appear), ask one short follow-up — count it within the two-question ceiling:

IA fidelity — (a) verbatim (same section sequence, same content beats; variants explore surface only: color, type, density, motion), or (b) reimagined (variants may demote / promote / drop sections, move IA priorities, take "what if" positions on the spine of the page). Default (b) for the typical refresh; pick (a) when the customer asked to keep their site structurally identical and only swap the surface.

If the user answers, stamp the chosen tier in direction.md § Movements as ia-fidelity: <tier>. If unanswered, default to reimagined and stamp ia-fidelity: reimagined (default).

Skip this question entirely when:

• The user's phrase already auto-pinned the axis (any trigger phrase in § 9 — "same IA, swap the surface" → verbatim; "what if we rethought the home page" → reimagined).
• An existing direction.md is being refined (the active tier holds unless the user explicitly re-pins).

The tier propagates to DESIGN.json.extensions.iaPriorities[].mutability: locked under verbatim, movable under reimagined. Phase 4 stamps the field; downstream prototype reads it.

Pair this question with the density-tuning question when both are unmoved — "two things to pin before we resolve: (1) density …, (2) IA fidelity …" — to stay within the two-question ceiling. If resolving register (brand vs product) is also outstanding, prioritise register first; defer one of density / ia-fidelity to the next turn.

Wait for the user's confirmation ("go", or a correction to the plan) before moving on.

Phase 2 — Resolve the divergence inputs

Once the plan is confirmed, resolve the divergence-toolkit inputs from skills/stardust/reference/divergence-toolkit.md. Before rolling the seed, run the mode-detection precedence below to decide whether the seed needs rolling at all.

Mode-detection precedence (run first)

The default mode for direct is determined by whether an extracted brand surface exists with usable signal — not by the user's freeform phrase alone. Stardust's primary use case is migrating an existing site with a design refresh; "make it modern" / "stunning new version" / "design fatigue cure" are migration-shaped asks. Treating those phrases as rebrand triggers (rolling a fresh divergence seed) produces output that is recognisably a different brand from the one that asked for the refresh — a published failure mode. The precedence below catches the common case as the default and reserves divergence-seed rolls for explicit rebrand requests.

The precedence is asymmetric on purpose: the safer mode (Mode A — brand-faithful) catches ambiguous phrases, and the riskier mode (rebrand / full divergence-seed) requires the user to name it.

1. Site migration / refresh — DEFAULT. When the captured brand signal stamped in § Setup step 6 is signal-strong, Mode A is active by default. The user's phrase moves expressive, distinctiveness, tone, and density axes inside Mode A — but palette and type are pinned to the captured surface, and the image-reuse contract (below) holds. Signature preservation also holds: a captured signature hero medium (background video / canvas / Lottie / scroll-motion, elevated as _brand-extraction.json#voice.heroMedium) or a site-wide motif is reproduced, not flattened — per skills/stardust/reference/ intent-dimensions.md § 8b, and exempt from the surprise budget. Surface in the plan: "Brand-faithful mode active — palette and type pinned to the captured brand surface. Pass --rebrand to override."

2. Rebrand — explicit opt-in. Mode A is overridden when any of the following triggers fire:

   • The user's phrase contains an explicit rebrand signal: any of rebrand, new brand, clean slate, start over, from scratch, replace the brand, not brand-faithful, editorial reimagination, completely new, redo the brand.
   • The --rebrand flag is passed.
   • Captured signal is signal-absent (no usable inheritance). Surface this case as an automatic switch with reason; the user can correct.

   In rebrand mode, run the standard divergence-seed roll per "Default mode (no constraints)" below. Surface in the plan: "Rebrand mode active — full divergence-seed roll. Mode A bypassed because ."

3. Brand-faithful + targeted exploration. When the user requests N variants (e.g. "3 variants", "4 directions") and Mode A is active, the variant role contract in § Multi-variant fork applies. Variant A is locked to strict Mode A (palette and type pinned, IA preserved, every improvements-list item applied). Variants B+ may amplify one captured trait (a motif, a photo treatment, an IA priority the current site underplays) but cannot:

   • introduce a font outside the captured surface,
   • introduce a color outside the captured palette,
   • shift the register from PRODUCT.md.

4. Signal-thin fallback. When captured signal is signal-thin, Mode A activates but warns: "Captured brand surface is thin — {reason}. Variants will inherit what is available; some dimensions will need to be filled by the divergence toolkit." The user may either re-run extract with a wider crawl (--cap 25) or proceed with reduced fidelity.

After mode-detection completes, run the existing mode definitions below (Mode A procedure, Mode B anchor-reference precedence, Mode C ground-family override) as applicable.

Mode A — Brand-faithful mode

Triggered automatically when the captured brand signal is signal-strong and no rebrand-mode override fired (see § Mode-detection precedence above), OR when the user pinned both type and palette (via explicit phrase: "keep typography and palette", "preserve the existing brand", "brand-faithful redesign"; or via constraints listing both as anchors).

In this mode, direct does not roll the type or palette dimensions of the seed — they are already locked. Going through the motions of font-deck and palette picks would be ceremony, producing picked_by = "user-constraint" records that don't reflect any real choice.

The mode procedure:

1. Record font_deck.name = "brand-inherited" and font_deck.picked_by = "user-constraint". Do not invoke reference/palette-picker.md.

2. Record palette.source = "inherited from _brand-extraction.json" and palette.picked_by = "user-constraint". Apply role-renaming per toolkit § 4 if the inherited names violate the brand-native rule (this is still useful — role renaming is presentational, not a divergence choice).

3. Still roll the seed for the non-locked dimensions (decade, register, ground-family-as-applicable per Mode C below). These dimensions still drive divergence — the visual register and the era can shift even when type and palette are pinned.

4. Auto-emit the brand_faithful_inversions[] block in extensions.divergence per reference/direction-format.md § Brand-faithful inversions. The list is mostly mechanical (see § Brand-faithful inversions in direction-format.md for the canonical patterns).

5. Surface in the user report which dimensions had teeth and which were inert:

   Divergence (brand-faithful mode):
     decade           ✓ rolled    → 2025-now
     craft            ✓ rolled    → Riso print
     register         ✓ rolled    → Memoir-adjacent
     ground-family    inherited   → stark-white (brand-native)
     font deck        inherited   → existing site stack
     palette          inherited   → existing 5-color set
   

6. Image-reuse contract. Captured images are reused via their public URLs (or the local copies in stardust/current/assets/media/ written by extract Phase 2) at the same semantic position as on the source site. Hero stays hero. Story-tile portrait stays story-tile portrait. Program-card image stays program-card image. Background-motif image stays background motif.

   This is part of brand-faithful inheritance, not a separate content rule. A variant that swaps a captured subject portrait for a gradient placeholder, or moves the captured hero photo to a card thumbnail, erases the brand's most load-bearing trust signal — the named-people stories that almost every nonprofit / service-led site has spent years building.

   The only legitimate ways to deviate from semantic position-preservation under Mode A:

   • The captured image is broken (404, blocked by referrer-policy, CORS-walled, or recorded with localPath: null and a downloadError).
   • The brand-review surfaced the image as a tension (e.g. T-stock-photography when added — flagging the captured image as obviously templated stock that the brand team would replace anyway).
   • The improvements list (Phase 2.5) explicitly notes a crop or positioning fix for that image — in which case the same image is reused at the corrected crop or position.

   Synthesised placeholders are forbidden under Mode A. When a captured image cannot be reused, the prototype shape brief declares the gap explicitly and the rendered prototype shows a placeholder-with-signature element so reviewers see the gap rather than a fabricated photo.

Mode A is the default whenever § Mode-detection precedence step 1 applies (captured signal is signal-strong and no rebrand override fires). It also activates explicitly when the resolved direction's constraints list contains brand-faithful AND explicit type AND palette anchors, OR when the user's phrase contains "keep typography" / "preserve the palette" / equivalent. The agent surfaces "Brand-faithful mode active" in the plan it shows the user before executing — the user can correct (e.g. "actually let me move the palette" or "actually rebrand it") before it locks.

Mode A+ — Brand-adjacent refinement (bounded, evidence-gated)

The middle tier between Mode A's hard pins and --rebrand. It exists because the median redesign candidate is a site whose brand is right but whose execution of that brand is part of the problem — a generic system body face, a palette whose only accent fails contrast on half its surfaces. Strict Mode A reproduces those weaknesses; rebrand throws away the brand. Mode A+ authorizes bounded upgrades, each gated on evidence:

• Same-classification type upgrade. When the improvements list (Phase 2.5) names the captured body or system face as a weakness with evidence (illegibility at captured sizes, a generic system stack where the brand deserves a voice, missing weights/axes the layout needs), the body face may be upgraded to a same-classification, deliverable face (humanist sans → humanist sans; grotesque → grotesque). The display face stays pinned — it carries the brand's recognition.
• Single-role palette recolor. When a specific captured color fails contrast or hierarchy as evidenced (computed ratio cited, or a T-color-imbalance tension), that one role may be re-derived (deepened, re-weighted) while every other role stays pinned. New hues from outside the captured family remain forbidden.

Contract: each refinement is recorded in DESIGN.json.extensions.divergence.brand_adjacent_refinements[] as { kind, captured, replacement, evidence, improvementsItem } — an inversion-style audit entry citing the improvements-list item that authorizes it. No evidence citation → no refinement. Mode A+ never activates by default: it requires the qualifying improvements-list item, and the plan surfaces each refinement explicitly ("body face upgraded Arial → Hanken Grotesk per improvements #3; display face pinned") so the user (or the hands-off record) sees exactly what moved. Refinements do not run through reference research — their justification is the captured weakness, not an external anchor.

Mode B — Anchor-reference precedence

When the user provides anchor references (Q1/Q2 answers like "Pentagram nonprofits, This American Life, NYT Opinion longform"), those references already imply seed dimensions. Pentagram implies decade 2025-now editorial. This American Life implies register Memoir-adjacent. Rolling those dimensions deterministically and getting an accidental alignment is fragile — the agent then has to retro-justify the alignment in direction.md.

Agent-sourced anchors (default when the user provides none). Mode B no longer waits for the user to name references: run the reference-research procedure (skills/stardust/reference/reference-research.md) to source 3–5 real-world anchors matched to the brand's category, register, and the resolved direction movement. Researched anchors carry the same implied-dimension weight as user-provided ones and are recorded as picked_by: "reasoned: <anchor>" with the full citation in extensions.divergence.references_used[]. User-provided references always outrank researched ones on conflict. When research is unavailable (ladder exhausted per reference-research.md § 1), Mode B degrades to the deterministic roll for the un-implied dimensions.

Precedence rule:

1. If anchor references are present, extract their implied dimensions:
   • Decade from era of the references (Pentagram → 2025-now; vintage Penguin → 1960s).
   • Craft from medium of the references (TAL → audio editorial ≠ a craft per se, but Bandcamp → web-print hybrid; Riso-print anthology → Riso).
   • Register from cultural reference set (Memoir, Tabloid, Catalogue, etc.).
   • Ground-family from typical ground of those references (NYT Opinion → cream/parchment; Pentagram nonprofit → stark-white or monochrome-tint).
2. Mark each implied dimension as picked_by = "anchor-reference: <ref-name>".
3. Roll the seed only for un-implied dimensions.
4. Record the anchor → dimension mapping in extensions.divergence.seed.anchors[].

Mode B can compose with Mode A: anchor-references narrow the seed, brand-faithful constraints lock type/palette, the remaining roll is whatever the anchors didn't already imply.

Mode C — Brand-faithful ground-family override

When Mode A is active and the seed's ground_family roll disagrees with the brand's existing ground (e.g. seed rolled monochrome-tint but the brand's captured background is #ffffff stark-white), the brand's ground wins. The seed roll is not discarded — it informs the alt-section surface instead (per divergence-toolkit.md § 4 Color roles). Record the override in extensions.divergence.seed.ground_family.override with one of three reasons:

• brand-faithful — Mode A active and brand has a fixed ground.
• print-paper — manual override for print/paper categories (existing toolkit rule).
• direction-driven — seed wins (default; no override).

The three reasons are mutually exclusive; surface the chosen one in the user report.

Default mode (no constraints)

When neither Mode A nor Mode B applies (rebrand or thin-signal runs with no user anchors), the procedure is research-first, roll as fallback:

• Reference research first. Run skills/stardust/reference/reference-research.md to source anchors for the brand's category and the resolved direction, and derive the dimensions they imply (picked_by: "reasoned: <basis>"). This is Mode B's machinery applied to agent-sourced anchors — see § Mode B above.
• Seed as fallback + tiebreaker. Roll the 4-dimension seed (decade × craft × register × ground-family) per § 2 of the toolkit only for dimensions research left un-implied, or entirely when research is unavailable. The roll also remains available as a deliberate convergence-breaker when the self-audit catches the model reproducing its defaults despite research. Record picked_by per dimension.
• Font deck. Pick from the 10 named decks per § 3, letting the researched anchors inform the pick the same way a seed implication would (e.g. an editorial-serif anchor set → serif-luxury). When neither research nor seed implies a deck, pick deterministically from the hash.
• Palette. If the resolved direction moves the color-energy axis or names the existing palette as part of the problem: derive a full role-ramped palette (text, grounds, borders, accents, states) from the primary researched anchor or a library candidate — the library (skills/direct/reference/palette-picker.md) is an anchor bank, not a closed menu; the model designs the ramp and validates every text-on-ground pair for WCAG AA before it lands in tokens (deterministic where it matters — math — not where it hurts — taste). Record the derivation basis in extensions.divergence.palette_source. Otherwise inherit the existing palette from stardust/current/_brand-extraction.json, applying role-renaming per toolkit § 4 if the inherited names violate the brand-native rule.

Always run

• Anti-toolbox audit. Regardless of mode, run the self-audit (toolkit § 1 Enforcement + Self-audit) on the resolved direction. Each anti-toolbox hit needs a brand-specific justification or it is removed.

Record every resolution in DESIGN.json.extensions.divergence per the v2 storage shape at the bottom of divergence-toolkit.md.

Phase 2.5 — Improvements list (Mode A only)

Before any variant is rendered downstream, write stardust/prototypes/<slug>-improvements.md listing 3–5 specific weaknesses observed in the captured site. This is the load-bearing artifact for variant A: without it, "make it better" has no claim the agent can defend, and each variant ends up inventing its own "better" — producing rebrand-shaped output even with Mode A active.

The improvements list is the brief variant A renders against. It is not prescriptive (it does not declare visual targets); it is descriptive of the gap between the existing site and a competent 2026 execution of the same brand. Variant A's job is to close the gap; variants B+ honor the list as a floor (they may go further but not contradict it).

Skip this phase when the resolved mode is rebrand — the improvements list assumes brand-faithful inheritance, and a rebrand replaces the site rather than fixing it.

Audit reuse. When stardust/audit/<domain-slug>/audit.json exists for this origin (written by stardust:audit), consume its design findings as candidate improvements instead of re-deriving from scratch — carry the finding IDs into each item's evidence citation. The specificity bar below still applies to every carried item.

What goes in the list

The list draws from five categories. Items should be specific enough that a downstream prototype shape brief can cite them by number.

1. Dated patterns the design world has moved past. Specific: "centered hero with stock photo + double CTA in primary blue is the SaaS template circa 2019" — not "the hero feels dated."
2. Cluttered IA, unclear hierarchy, weak CTAs, redundant sections. E.g. "home page has 4 different donor CTAs, each with a different verb (DONATE / GIVE / SUPPORT / CONTRIBUTE), fragmenting the conversion funnel."
3. Contrast failures, accessibility gaps, density issues. Pull from brand-review.html Tensions when present (T-color-imbalance, T-img-alt-empty, etc.). E.g. "Primary CTA #008192 on white passes AA at 4.6:1 but the same teal on light-grey card surfaces drops to 3.1:1 — fails AA on those instances."
4. Cliché conventions the brand could move past while staying recognisably itself. E.g. "All headings render uppercase via CSS — the brand voice survives mixed-case headlines, and mixed-case reads as more current without changing identity."
5. Missed opportunities the existing site doesn't capitalise on. E.g. "The captured photography of named program participants is excellent but the home page renders all 6 portraits as 280×180 thumbnails in a slick-slider; the photographs support full-bleed editorial treatment that would carry the trust signal far more effectively."

Specificity bar

A weakness is specific enough when it cites:

• a measurable observation (size, ratio, contrast value, count, or named tension ID) drawn from the per-page JSON, the brand-review, or the brand-extraction;
• the design pattern at fault (named, e.g. "centered hero + dual CTA");
• one concrete fix the variant A brief will apply.

"The hero needs work" fails all three. "Hero photo is cropped to 280×180 in a 1440-wide viewport when the captured source supports 16:9 full-bleed at 1440×810; variant A fix: render at full-bleed and move the headline to a left-anchored two-column overlay" passes all three.

Format

Markdown, with a _provenance frontmatter block per the artifact-map convention. Each item is a numbered list entry with a category tag, a weakness statement, and a one-line fix. Example:

<!--
_provenance:
  writtenBy: stardust:direct
  writtenAt: 2026-04-29T11:00:00Z
  readArtifacts:
    - stardust/current/_brand-extraction.json
    - stardust/current/brand-review.html
    - stardust/current/pages/<slug>.json
  stardustVersion: 0.10.x
-->

# Improvements — <slug>

1. **[dated-pattern]** Centered hero with double CTA pair (DONATE +
   LEARN MORE) is the 2019 nonprofit-template silhouette.
   *Fix:* Replace with a left-anchored editorial composition; one
   primary CTA, one secondary text-link.

2. **[ia-clutter]** 4 distinct donor verbs across the home page
   (DONATE, GIVE, SUPPORT, CONTRIBUTE) fragment the funnel.
   *Fix:* Pick one canonical verb (DONATE, per the CTA frequency
   table); other instances become secondary "see all ways to give"
   links.

3. **[contrast]** Brand teal on light-grey card surfaces resolves to
   3.1:1 — fails WCAG AA. (See `T-color-imbalance` in brand-review.)
   *Fix:* Reserve teal for white-ground only; use deepened teal
   (#005a68) on grey surfaces.

4. **[cliché]** All headings render uppercase via CSS, including
   long-form section openers ("OFFICIAL FOUR STAR CHARITY"). The
   shout reads as urgency at first heading and as fatigue by the
   third.
   *Fix:* Mixed-case for headings ≥3 words; preserve uppercase only
   for short imperative CTAs and eyebrow labels.

5. **[missed-opportunity]** Six named-participant portraits render
   as 280×180 thumbnails in a slick-slider; the captured source
   supports editorial-scale treatment.
   *Fix:* Replace carousel with a 3-column grid; portraits at 4:5
   aspect, 1:1 minimum 480px wide.

Stopping condition

If after reading the brand-review, the per-page JSON, and the brand-extraction the agent cannot name 3 specific weaknesses that meet the specificity bar, stop. Variant A has no brief; the "better" claim fails. See § Failure modes (d).

The agent should not rationalise — "the hero is dated" is not an item, "the typography could be more modern" is not an item. Genuine empty-list cases occur when the captured site is already at a high execution level on observable dimensions. In that case the honest answer is to surface the empty list and propose reduced scope (density-and-contrast adjustments only, or pivot to a single exploratory variant rather than three).

Phase 2.6 — Multi-variant fork (when N > 1)

When the user requests N variants (phrase contains "3 variants", "4 directions", or equivalent), variant slots are role-differentiated, not seed-differentiated. Each slot serves a distinct decision the customer is making — variants exist to de-risk a brand decision, not to fan out into N rebrand explorations.

The previous default behavior (each variant rolls a fresh divergence seed with anchor references picked per slot) produced what the internal review called "three rebrands" output: each variant was defensible standalone but none felt like the same brand. The variant role contract below addresses this directly.

Branch on ia-fidelity first

The variant role contract is ia-fidelity-aware. Read the tier stamped in Phase 1 (per reference/intent-dimensions.md § 9 and the IA-fidelity tuning step above), and branch:

Under ia-fidelity: verbatim — surface-tuning forks (A1/A2/A3).

Variant slots are A1 / A2 / A3 (or A1…An for N > 3), all surface-tuning forks of A's role. Each must differ from the others by ≥ 2 surface changes drawn from:

• type-weight choice (e.g. 400 vs 600 vs 800 display)
• type-scale ratio (e.g. 1.2 vs 1.333 vs 1.5)
• density tier (within the multi-audience floor in § 4)
• motion energy (still vs gentle vs animated)
• color-temperature move within the captured palette (warm-leaning vs cool-leaning vs neutral-balanced)
• spacing rhythm (compact vs even vs generous within the floor)

Forbidden differentiation axes under verbatim:

• section sequence
• section presence / absence
• IA priority
• layout strategy of a major section

A1 / A2 / A3 names indicate "different tunings of the same role," not different roles. When the user asks for "3 variants" while ia-fidelity is verbatim, the fork produces A1 / A2 / A3 automatically; for N=1, just A.

The improvements list (Phase 2.5) still anchors variant A's role — A1/A2/A3 all apply every item from <slug>-improvements.md. They differ only in surface treatment of the same applied fixes.

The convergence detector in prototype/SKILL.md Discipline 10 becomes its inverse under verbatim: structural deltas (section-sequence / presence / IA priority / layout strategy) are forbidden, surface-only deltas are required.

Under ia-fidelity: reimagined — A + B + C role-differentiated.

Variant slots are A + B + C (or A + B + C + D…) per the variant role contract below. The contract is unchanged from the prior behavior: faithful + improvements / one captured trait amplified / different captured trait amplified.

The remainder of this Phase 2.6 (variant role contract, variant differentiation contract, C-cliff failure mode) applies as written under reimagined. Under verbatim, only the surface-fork rules above apply; the variant role contract below does not.

Variant role contract

Variants beyond C (D, E, …) follow the C+ contract — each amplifies a distinct captured trait, declared in the shape brief.

Surface forks of role-differentiated variants (B1/B2/B3, C1/C2/C3…)

Under ia-fidelity: reimagined, a variant's role (A, B, C, …) may spawn surface forks — tunings of the same role across the same surface axes that A1/A2/A3 use under verbatim. The pattern:

• B is "scroll cinema amplified" (the role).
• B1, B2, B3 are surface tunings of B that all amplify scroll cinema, but vary along type-weight, type-scale, density, motion-energy, color-temperature, or spacing-rhythm.

This composes the verbatim-mode surface-fork machinery with the reimagined-mode role differentiation: the role contract binds the captured trait being amplified; the surface-fork delta governs how that trait reads in chrome.

Surface forks of role-differentiated variants are opt-in, not default. The default fork under reimagined is still A + B + C. Surface forks appear in two ways:

1. Explicit user request: "design 5 variant directions for B" — the user asks for surface tunings of an existing role- differentiated variant. Render B1 / B2 / B3 / B4 / B5, each amplifying B's captured trait but with distinct surface tunings.
2. Via --add-variant <name> with a parent declared: e.g., --add-variant B3 after B exists. The parent inferred from the slot name's letter prefix (B3 → parent B). See § Add-variant mode → Per-field inheritance rules for the inheritance chain.

Surface forks of role-differentiated variants follow the same ≥ 2 surface-changes contract as A1/A2/A3 — the per-pair differentiation is surface-only (type-weight / type-scale / density / motion-energy / color-temperature / spacing-rhythm), with the captured trait being amplified held constant across the fork. Structural differentiation (section sequence / presence / IA priority / layout strategy) is forbidden between B and its surface forks B1/B2/B3 — the role IS the captured trait, and changing structure changes the role.

The variant-convergence detector in prototype/SKILL.md Discipline 10 reads the variant's parent slot: when comparing B vs B3 (parent–child surface fork), it applies the verbatim-style inverse rule (surface deltas required, structural deltas forbidden); when comparing B vs C (sibling role-differentiated variants), it applies the reimagined-style rule (structural deltas required).

The cap on motion-energy and other amplified surface axes is the parent role's cap — a B3 that amplifies motion-energy beyond B's ceiling must declare a cap override in its DESIGN.json extensions with a one-sentence rationale (e.g. "B3 raises B's motion choreography count from ≤ 3 to ≤ 5 to accommodate the loud- register tuning of scroll cinema"). The override must cite a captured-source basis or it refuses; cap overrides without a captured-source rationale produce surface drift instead of trait amplification.

Variant differentiation contract

Each pair of variants must differ by ≥ 2 substantive changes drawn from this set:

• section sequence (which sections appear in which order),
• section presence / absence (a section in one variant, not in another),
• layout strategy of a major section (e.g. hero is split-half vs. full-bleed-photo vs. type-led),
• IA priority (which audience leads the home — donor vs. recipient vs. volunteer for nonprofits; product vs. story for commerce).

Variants that fail the ≥ 2 changes test are rendered as the same variant under different chrome — "variants are barely different" is the published failure mode and is grounds for refusing render. See § Failure modes (b) — when only 1–2 captured traits are distinct enough to amplify, the agent surfaces this and proposes 1 or 2 variants rather than producing 3 weak ones.

The C-cliff failure mode

When defin

…(truncated)