Krea-generate skill improvements

[alcybiades]

• Make the model more inclined to use edit models when the user is implicitly asking for one
• consolidate images.md into SKILL.md
• improve the arch viz skill

[github-actions]

Claude finished @alcybiades's task in 3m 9s —— View job

---

PR Review: Krea-generate skill improvements

• Read the full diff against origin/main
• Review edit-model triggering changes in SKILL.md
• Verify image.md consolidation (no broken references left behind)
• Review archviz workflow + eval changes
• Check cross-references, model names, frontmatter/version integrity
• Post inline comments + summary

Overall this is a solid, well-targeted change. The image.md → SKILL.md consolidation is clean — I grepped the whole repo and every workflows/image.md reference (README, krea-marketing/references/artifact-taxonomy.md, key-visual-sheet.md, lora-train-and-use.md, image-text-poster.md) was repointed to the new Image Workflow, with no danglers left behind. Version-sync sites are untouched and still all read 0.6.1, so validate.yml won't break. The implicit-edit-request guidance is a genuine UX improvement and clearly written.

Three findings, posted inline:

1. SKILL.md:80–82 — duplicate list numbering. The MUST-do checklist is 1., 2., 2. instead of 1./2./3.. Minor, but this is a hard-prescription list and mis-numbering reads like a merged/optional step. Suggestion included. Fix →

2. enhance.md:22–23 (and 45–46) — model IDs have a stray leading slash. /topaz/bloom-enhance / /topaz/standard-enhance are inconsistent with every other ID in the repo (topaz/standard-enhance, google/nano-banana-pro, krea/krea-2/large). Since the skill tells the agent to match by live id, a mis-formatted literal invites a bad exact-match. Suggestion included. Fix →

3. archviz-3d-to-render.md — 4K-as-default vs. cost-preflight and the evals. Defaulting nano-banana-pro to 4K for every render means the cost-preflight (a hard "wait for explicit yes") now fires before the first structural pass, removing the old cheap 1K-first-pass path. The workflow's own "Cost & time" section (lines 47–48) still describes 1K drafts as normal and is now stale, and HC-12 still expects the first 4K render to proceed ungated (only generate_video is gated). Grading softens that to a WARNING so nothing hard-fails, but the skill and eval now describe different behavior. Worth reconciling. Details →

None of these block merge; #3 is the one worth a decision (update the docs/eval to embrace the 4K default, or keep a cheap 1K first pass).

[github-actions]

Duplicate list marker: this ordered list is 1., 2., 2. — the "Feed the existing image…" step should be 3.. Skill prose is the product, and a mis-numbered MUST-do checklist reads as if one step is optional/merged. Renumber:

Suggested change

	1. Use an editing/reference-capable model such as `google/nano-banana-pro`, or `openai/gpt-image-2` when the request is complex, premium, or text-heavy.
	2. Scan the conversation and track down the prior (or relevant) output and its associated prompt.
	2. Feed the existing image output into the model as a conditional image input using the exact reference/source/image field from the live schema.
	1. Use an editing/reference-capable model such as `google/nano-banana-pro`, or `openai/gpt-image-2` when the request is complex, premium, or text-heavy.
	2. Scan the conversation and track down the prior (or relevant) output and its associated prompt.
	3. Feed the existing image output into the model as a conditional image input using the exact reference/source/image field from the live schema.

[github-actions]

Model-ID formatting is inconsistent with the rest of the repo. Everywhere else IDs are written without a leading slash (topaz/standard-enhance, google/nano-banana-pro, krea/krea-2/large, bytedance/seedance-2). Here (and again at lines 45–46) they're written /topaz/bloom-enhance and /topaz/standard-enhance. Since the skill tells the agent to "match by live id/name/description," a mis-formatted literal id invites a bad exact-match attempt. Drop the leading slash:

Suggested change

	| Creative enhance | Topaz Bloom, `/topaz/bloom-enhance` | Default for creative detail injection, relight, polish, or premium-looking restyles. |
	| Faithful / realistic enhance | Topaz Standard, `/topaz/standard-enhance` | Default for normal faithful enhancement, upscaling, sharpening, denoise, cleanup, and photography. |
	| Creative enhance | Topaz Bloom, `topaz/bloom-enhance` | Default for creative detail injection, relight, polish, or premium-looking restyles. |
	| Faithful / realistic enhance | Topaz Standard, `topaz/standard-enhance` | Default for normal faithful enhancement, upscaling, sharpening, denoise, cleanup, and photography. |

Also apply the same fix to the two IDs in the Recipe (lines 45–46).

[github-actions]

Making google/nano-banana-pro at 4K the default for every archviz render (step 2) has two knock-on effects worth confirming were intended:

• Step 5 requires a cost-preflight for 4K, and per ../references/cost-preflight.md a 4K render is a hard "show estimate → wait for explicit yes" gate. So the workflow now stops for approval before the first structural render — a change from the old cheap 1K-first-pass model. The unchanged Cost & time section (lines 47–48) still frames "draft 1K is moderate" as the normal path, so it's now stale; update it to reflect the 4K default.
• evals/hero/cases/HC-12-archviz-walkthrough-chain.json still expects the first 4K render to proceed ungated (gated_steps: ["generate_video"], approval_turn: 3 — only the video is gated). That contradicts a 4K-always default that trips cost-preflight before rendering. Grading softens it to a WARNING so it won't hard-fail, but the skill and eval now describe different behavior — reconcile them.

Either update Cost & time (and HC-12) to match the 4K default, or keep a 1K structural first pass as the default and reserve 4K for the accepted final.

[albertsalgueda]

lgtm