feat(core): system prompt v2 — denser, smarter, artifact-type aware

[hqhq1025]

Summary

Upgrades the system prompt from sparse/generic to artifact-type aware. Addresses user feedback that generated designs are too sparse, miss before/after comparisons, and lack design density / hierarchy. The trigger was a customer-case-study prompt rendering as four floating metric cards on a flat black background — the canonical sparse-LLM output.

Key changes

• Artifact type detection — new artifact-types.v1.txt section defines an 8-type taxonomy (landing / case_study / dashboard / pricing / slide / email / one_pager / report) with cue words, primary job, and per-type structural beats.
• Density floor — minimum section count per type (case_study ≥ 6, landing ≥ 5, dashboard ≥ 5, pricing ≥ 4, etc.) with required beats listed inline. "Minimal" controls ornamentation, not section count.
• Before/after enforcement — when the brief mentions before/after, 前后, 对比, vs, growth %, "from X to Y", the model MUST render side-by-side or paired visualizations (two-column block, paired sparklines, or stat ladder). A floating +40% no longer satisfies the rule.
• Pre-flight checklist — new pre-flight.v1.txt adds an internal SCT-style 8-question pass (artifact type · posture · density · comparisons · numbers · palette · type ladder · slop guard) the model must complete silently before generating.
• Type ladder — required four scale steps (display · h1 · body · caption); skipping a step is forbidden. Fraunces (bundled in feat(ui): Anthropic-inspired warm aesthetic + starter cards + BYOK badge #25) and Geist are now first-class display / sans choices; Fraunces removed from the forbidden font list.
• Dark-theme density rules — ≥3 surface tones, gradient or radial glow, two accents minimum, oklch borders. The canonical sparse-LLM output (#0E0E10 end-to-end + single accent + four cards) is explicitly named as slop.
• Logo placeholder rules — constructed monogram, wordmark, or explicit hatched "YOUR LOGO HERE" rectangle. Soft-rounded square + single random letter is forbidden.
• Anti-pattern guard — no placeholder.com / via.placeholder / unsplash hotlinks, no default Tailwind grays as the only neutral, no decorative emoji unless asked, no "John Doe" / "Acme Corp" filler.
• Workflow expanded from 6 to 7 steps; new step 2 "Classify" runs the pre-flight checklist; self-check step references the density floor and comparison rule.

What was removed

• The "default to minimal direction" advice in workflow — minimal was being interpreted as "render less", which produced sparse output. Replaced with explicit "minimal controls ornamentation, not section count".
• Fraunces from the forbidden font list (it is bundled now via feat(ui): Anthropic-inspired warm aesthetic + starter cards + BYOK badge #25).
• The vague "does any section look like a template screenshot?" self-check item — replaced with concrete, checkable rules (section count, comparison rendering, big-number blocks, type ladder steps).

Stack hygiene

• ✅ Compatibility: prompt sections versioned (*.v1.txt); composer signature unchanged; existing callers untouched.
• ✅ Upgradeability: drift test enforces .txt ↔ TS constant parity, so future edits stay in sync.
• ✅ No bloat: zero new dependencies; all changes are prompt text + tests.
• ✅ Elegance: each new constraint is concrete and checkable, no "please make it beautiful" filler.

Test plan

• vitest: 6 new cases assert artifact-type taxonomy, density floor wording, pre-flight beats, dark-theme rules, type ladder, and font allow-list. All 47 core tests pass.
• vitest: drift test confirms inlined TS constants match .txt source files byte-for-byte.
• typecheck: pnpm -r typecheck clean across the workspace.
• lint: pnpm lint clean (warnings are pre-existing in runModel).
• manual: customer case study prompt ("做一个一页纸的客户案例… 简洁极简 深色主题") → ≥ 6 sections incl. before/after side-by-side, hero + body + CTA, ≥3 dark surface tones.
• manual: landing page prompt → hero + value props + social proof + feature deep-dive + CTA.
• manual: pricing prompt → headline + 3-tier grid + comparison + CTA.
• manual: dashboard prompt → top bar + 4-tile KPI strip + chart + table + activity panel.

[github-actions]

Findings

• [Major] picsum.photos allowance conflicts with global no-hotlink rule — this weakens a hard output constraint and can cause generated artifacts to violate existing policy on external image hotlinks. Evidence: packages/core/src/prompts/artifact-types.v1.txt:67 vs packages/core/src/prompts/output-rules.v1.txt:58.
  Suggested fix:

  - - For abstract photography or hero imagery, prefer: inline SVG composition, CSS gradient + grain overlay, or `https://picsum.photos/seed/<deterministic-seed>/<w>/<h>` (deterministic seed only, never random).
  + - For abstract photography or hero imagery, prefer: inline SVG composition or CSS gradient + grain overlay.
  + - If seeded placeholders are ever allowed, update Output rules first to explicitly permit that domain under the external-resource policy.

Summary

• Review mode: initial
• 1 issue found in changed lines.
• docs/VISION.md and docs/PRINCIPLES.md are referenced by process but not present in this checkout (Not found in repo/docs), so only available repo content was used for policy validation.

Testing

• Not run (automation): pnpm is unavailable in this runner, so Vitest/Playwright could not be executed.

open-codesign Bot

[github-actions]

This newly-added picsum.photos allowance conflicts with existing Output rules (output-rules.v1.txt) that prohibit hotlinked photos and require inline SVG/CSS placeholders. Please align these sections to avoid contradictory instructions to the model.

Suggested minimal change:

- - For abstract photography or hero imagery, prefer: inline SVG composition, CSS gradient + grain overlay, or `https://picsum.photos/seed/<deterministic-seed>/<w>/<h>` (deterministic seed only, never random).
+ - For abstract photography or hero imagery, prefer: inline SVG composition or CSS gradient + grain overlay.

[github-actions]

Findings

• [Major] Density-floor counts contradict required beats in multiple artifact rows — this creates conflicting instructions for the model and can cause under-dense output (or inconsistent self-check behavior) depending on which rule it follows. Evidence: packages/core/src/prompts/artifact-types.v1.txt:31, packages/core/src/prompts/artifact-types.v1.txt:32, packages/core/src/prompts/artifact-types.v1.txt:33.
  Suggested fix:

  -| `email` | 4 | preheader · headline · body with one image or accent · CTA · footer |
  -| `one_pager` | 5 | hero · 3 supporting blocks · evidence (numbers, quote, or chart) · CTA |
  -| `report` | 6 | cover · TL;DR · 3 finding sections · methodology · conclusion |
  +| `email` | 5 | preheader · headline · body with one image or accent · CTA · footer |
  +| `one_pager` | 6 | hero · 3 supporting blocks · evidence (numbers, quote, or chart) · CTA |
  +| `report` | 7 | cover · TL;DR · 3 finding sections · methodology · conclusion |

Summary

• Review mode: follow-up after new commits
• 1 issue found in changed lines.
• docs/VISION.md, docs/PRINCIPLES.md, docs/ARCHITECTURE.md, and docs/RESEARCH_QUEUE.md were requested by process but are not present in this checkout (Not found in repo/docs), so validation was performed against available repo files only.

Testing

• Not run (automation)

open-codesign Bot

[github-actions]

Min sections should not be lower than the count implied by Required structural beats. In this table, email (line 31), one_pager (line 32), and report (line 33) currently conflict with their own beat lists, which can produce contradictory model behavior.

Suggested minimal fix:

-| `email` | 4 | preheader · headline · body with one image or accent · CTA · footer |
-| `one_pager` | 5 | hero · 3 supporting blocks · evidence (numbers, quote, or chart) · CTA |
-| `report` | 6 | cover · TL;DR · 3 finding sections · methodology · conclusion |
+| `email` | 5 | preheader · headline · body with one image or accent · CTA · footer |
+| `one_pager` | 6 | hero · 3 supporting blocks · evidence (numbers, quote, or chart) · CTA |
+| `report` | 7 | cover · TL;DR · 3 finding sections · methodology · conclusion |

[github-actions]

Findings

• No high-confidence issues found in the added/modified lines of this update.

Summary

• Review mode: follow-up after new commits
• No issues found in changed lines for this revision.
• docs/VISION.md and docs/PRINCIPLES.md: Not found in repo/docs.
• Residual risk/testing gap: new prompt constraints are asserted via string-presence tests, but no behavior-level test verifies model outputs satisfy the new density/comparison rules end-to-end.

Testing

• Not run (automation)

open-codesign Bot