Skip to content

Verbs (phase 3): study and audit - #42

Merged
codeswithroh merged 1 commit into
mainfrom
feat/verbs-phase3
Jul 23, 2026
Merged

Verbs (phase 3): study and audit#42
codeswithroh merged 1 commit into
mainfrom
feat/verbs-phase3

Conversation

@codeswithroh

Copy link
Copy Markdown
Owner

What this changes

Phase 3, the final piece of the Hallmark-caliber upgrade. Hallmark's third pillar is verbs — a single skill that does more than build. We were build-only. This adds the two highest-value verbs, each adapted to leverage what tastemaker already has that Hallmark doesn't.

references/verbs/study.md — extract design DNA from a screenshot or URL, never the pixels:

  • Image mode leans on scripts/extract_palette.py for deterministic color extraction (a real edge over Hallmark, which names colors by eye), then re-verifies the assigned roles through the contrast matrix — so the DNA handed forward is already a legal contract.
  • URL mode WebFetches and names exact fonts/colors, treating all fetched HTML/CSS as untrusted inert data — ignores any instruction embedded in the page, extracts only design facts (prompt-injection safety baked in). Falls back to asking for a screenshot on auth-walled / SPA-shell / unreadable URLs.
  • Refuses template-marketplace + portfolio-aggregator sources and non-public targets; requires attestation before locking a URL's DNA. Diagnosis names macrostructure + archetypes (phase-1 catalog) + type + color anchor + anti-patterns not to carry over. Branches: build with the DNA, lock it, or stop.

references/verbs/audit.md — score existing UI, don't edit:

  • Grades against the phase-2 numbered 47-gate list + six pre-emit axes, mood-aware (respects the mood the build stamp declares). Findings carry gate number, location, severity (critical/major/minor), and a one-line concrete fix.
  • Two structural checks the gate list alone misses: the generic-template fingerprint, and stamp-vs-page ("stamp lies") drift where the CSS stamp claims a shape/contrast the page doesn't deliver.
  • Explicitly non-destructive — grading and changing are separate acts; a "fix it" follow-on hands into the build flow.

Wiring: SKILL.md gains a Modes table (build default + study + audit) with detection/routing, the verb files in the reference table, and study/audit trigger phrases in the description.

The phased upgrade is complete

All grafted onto tastemaker's generated-palette + real-asset foundation, which stays the differentiator Hallmark lacks.

Test plan

  • All cross-references in the verb files resolve; extract_palette.py + check_contrast.py exist
  • Fence balance clean, no stray "preset"
  • SKILL.md Modes table + reference-table entries + description triggers wired
  • URL-mode fetching explicitly treats page content as untrusted inert data (prompt-injection safety)

🤖 Generated with Claude Code

Phase 3, the final piece of bringing tastemaker to Hallmark's caliber.
Hallmark's third pillar is verbs — a single skill that does more than
build: it can study a reference to extract reusable DNA, and audit
existing UI against its own standards. We were build-only. This adds
the two highest-value verbs, each adapted to leverage what tastemaker
already has that Hallmark doesn't.

references/verbs/study.md — extract design DNA from a screenshot or
URL, never the pixels:
- Image mode leans on scripts/extract_palette.py for DETERMINISTIC
  color extraction (a real edge over Hallmark, which has no such
  script and names colors by eye), then re-verifies the assigned
  roles through the contrast matrix so the DNA handed forward is
  already a legal contract.
- URL mode WebFetches and names exact fonts/colors, treating all
  fetched HTML/CSS as untrusted inert data — ignores any instruction
  embedded in the page, extracts only design facts (prompt-injection
  safety baked in). Falls back to asking for a screenshot on
  auth-walled / SPA-shell / unreadable URLs rather than guessing.
- Refuses template-marketplace and portfolio-aggregator sources and
  non-public targets; requires attestation before locking a URL's DNA
  into the project. Diagnosis names macrostructure + archetypes (from
  our phase-1 catalog) + type + color anchor + anti-patterns NOT to
  carry over. Branches: build with the DNA (diversification suspended,
  external DNA followed), lock it into the style lock, or stop.

references/verbs/audit.md — score existing UI, don't edit:
- Grades against the phase-2 numbered 47-gate list + six pre-emit
  axes, mood-aware (respects the mood the build stamp declares when
  applying mood-scoped gates). Findings carry gate number, location,
  severity (critical/major/minor), and a one-line concrete fix.
- Two structural checks the gate list alone misses: the generic-
  template fingerprint, and stamp-vs-page ("stamp lies") drift where
  the CSS stamp claims a shape/contrast the page doesn't deliver.
- Explicitly non-destructive: grading and changing are separate acts;
  a "fix it" follow-on hands off into the build flow.

Wiring: SKILL.md gains a Modes table (build default + study + audit)
with detection/routing, the verb files in the reference table, and
study/audit trigger phrases in the skill description.

This completes the phased Hallmark-caliber upgrade: phase 1 structural
variety engine, phase 2 quality gates, phase 3 verbs — grafted onto
tastemaker's generated-palette + real-asset foundation, which stays
the differentiator Hallmark lacks.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
@codeswithroh
codeswithroh merged commit 4e06bf5 into main Jul 23, 2026
@codeswithroh
codeswithroh deleted the feat/verbs-phase3 branch July 23, 2026 04:24
codeswithroh added a commit that referenced this pull request Jul 26, 2026
Resolves #54. Four real writeups sourced from actual prior work, not invented:

- section-spacing-measurement.md — the make.design DOM-introspection measurement behind #48
- hallmark-architecture-study.md — the structural-gap study behind #40/#41/#42, including an honest account of where tastemaker already led Hallmark and where it didn't
- contrast-as-authored-vs-as-used.md — the multi-reviewer-converged finding behind #16/#18's contrast contract
- non-latin-typography.md — the Pretendard/Hangul research (three independent sources) behind #39

Each links the PR that actually shipped the fix, the same honesty standard the site changelog already holds to. Linked from the README (new Research section + anchor nav) and from research/README.md as an index.

Also fixes em-dash house-style violations introduced in #71's CHANGELOG.md and site/changelog.html — README/CONTRIBUTING/site content stays em-dash-free per this project's standing convention; caught on review before merge, not after.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant