Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Render an interactive, self-contained HTML companion for a GEO content brief (04-content-brief) or a publish-ready draft (05-production), so a NON-technical client reviewer (founder, organizer staff, the domain expert filling slots) can fill REQUIRED-FILL slots, leave section-level comments, and approve/return work in the browser instead of editing Markdown. Use when a brief or draft needs to go to a client/expert for review, or when building the briefs/index.html entry page for a client folder.
.claude/skills/viryazheng-geo-review-html/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | 571% | 0% |
| case-02 | ✗→✓ | ▲ Improved | 402% | 0% |
| case-04 | ✗→✓ | ▲ Improved | 0% | 0% |
| case-06 | ✗→✓ | ▲ Improved | -14% | 0% |
| case-07 | ✗→✓ | ▲ Improved | -3% | 0% |
The Markdown briefs/drafts are right for the GEO operator; they are awkward for the client side. This skill renders a companion HTML alongside the Markdown so non-technical reviewers can work in the browser. Battle-tested in the SEA-CICSIC 9-draft batch (see issue #1).
Division of labour (do not blur this):
templates/review.html owns localStorageauto-save, JSON export, the progress counter, comment toggles, and the approve/return state machine. It is correct as written; do not regenerate it per run.
frontend-design skill applies. When a client wants a branded or more distinctive look, restyle the <style> block of templates/review.html using frontend-design principles. --accent is already overridable per client via brand_context (see below); deeper restyling edits the CSS.
render_html.py only parses Markdown + meta into a payload and injects it.04-content-brief Step 7 — emit briefs/<id>.html (fill-form) next to the .md.04-content-brief Step 9 — ingest the reviewer's returned *.feedback.json.05-production Step 7 — emit drafts/<id>.html (review form) next to the .md.briefs/index.html entry page needs (re)building.bashSK=plugins/recomby-geo/skills/geo-review-html # Flavor A — fill-form HTML for a brief with REQUIRED-FILL slots python3 $SK/scripts/render_html.py --mode brief \ --md clients/<slug>/briefs/<id>.md \ --meta clients/<slug>/briefs/<id>.meta.json \ --brand clients/<slug>/brand_context.json \ --out clients/<slug>/briefs/<id>.html # Flavor B — review HTML for a publish-ready draft (section comments + approve) python3 $SK/scripts/render_html.py --mode draft \ --md clients/<slug>/drafts/<id>.md \ --meta clients/<slug>/drafts/<id>.meta.json \ --brand clients/<slug>/brand_context.json \ --out clients/<slug>/drafts/<id>.html # Entry page — one card per brief/draft, live progress, status tags python3 $SK/scripts/render_html.py --mode index \ --client-dir clients/<slug> \ --brand clients/<slug>/brand_context.json \ --out clients/<slug>/briefs/index.html
Pure standard library — no install step. Output is a single self-contained .html (CSS + JS inlined) safe to email to a client; they double-click it.
REQUIRED-FILL · <id> · <type>blockquote blocks. Each becomes a yellow textarea with a <details> dropdown (what / why / bad example / good example). Everything else becomes a green "already-filled" prose block with its own 💬 comment toggle. The authoritative slot list is meta.json's slots[].
## headings so every section gets a commenttoggle; the prose is rendered read-only with an ✓ Approve as-is action.
briefs/ and drafts/ for *.meta.json, readingformat, status, word_count, and slot fill counts. Live per-brief progress is read from each brief's localStorage when the page is opened.
--brand is optional. brand_name sets the client label; brand.primary_color(or brand_color) overrides the --accent CSS variable for light branding.
Each HTML exports a JSON file matching schemas/review_feedback.schema.json:
json{ "brief_id": "free-cash-prize", "filled_at": "2026-06-07T14:30:00Z", "filled_by": "Jane Doe", "answers": { "data-1": "...", "quote-1": "..." }, "comments": { "overall": "...", "filled-headline-finding": "..." }, "status": "reviewed-with-comments" }
status ∈ reviewed-with-comments | approved-as-is | partial-fill. The GEO operator feeds answers into the matching slots during 04-content-brief Step 9 (NEVER auto-filled by AI — the slots are the moat).
answers — the JSON carries the human's real input;that is the whole point. AI assembles, humans supply.
but don't rewrite the localStorage/export/state-machine logic per run.
non-technical reviewer can open it from an email attachment.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→pass | 4,792 | 2,401 | -50% | 1 | 1 | 0% | 269 | 1,805 | +571% | 0 | 0 | — |
case-02 | fail→pass | 5,503 | 2,383 | -57% | 1 | 1 | 0% | 355 | 1,781 | +402% | 0 | 0 | — |
case-03 | fail→fail | 4,516 | 6,449 | +43% | 1 | 1 | 0% | 200 | 1,732 | +766% | 0 | 0 | — |
case-04 | fail→pass | 15,885 | 6,670 | -58% | 1 | 1 | 0% | 2,356 | 2,357 | +0% | 0 | 0 | — |
case-05 | pass→fail | 15,679 | 5,625 | -64% | 1 | 1 | 0% | 2,495 | 2,251 | -10% | 0 | 0 | — |
case-06 | fail→pass | 13,784 | 3,856 | -72% | 1 | 1 | 0% | 2,236 | 1,916 | -14% | 0 | 0 | — |
case-07 | fail→pass | 13,143 | 4,379 | -67% | 1 | 1 | 0% | 2,088 | 2,027 | -3% | 0 | 0 | — |
case-08 | fail→pass | 8,821 | 3,737 | -58% | 1 | 1 | 0% | 1,503 | 1,879 | +25% | 0 | 0 | — |
case-09 | fail→pass | 13,601 | 5,205 | -62% | 1 | 1 | 0% | 2,441 | 2,315 | -5% | 0 | 0 | — |
case-10 | pass→pass | 9,617 | 3,681 | -62% | 1 | 1 | 0% | 1,384 | 1,900 | +37% | 0 | 0 | — |
case-11 | fail→pass | 12,497 | 4,388 | -65% | 1 | 1 | 0% | 1,814 | 1,967 | +8% | 0 | 0 | — |
case-12 | fail→fail | 13,177 | 6,160 | -53% | 1 | 1 | 0% | 2,051 | 2,215 | +8% | 0 | 0 | — |
case-13 | pass→pass | 7,923 | 2,085 | -74% | 1 | 1 | 0% | 1,249 | 1,712 | +37% | 0 | 0 | — |
case-14 | fail→pass | 6,572 | 9,586 | +46% | 1 | 1 | 0% | 1,013 | 3,034 | +200% | 0 | 0 | — |
case-15 | fail→pass | 6,591 | 3,433 | -48% | 1 | 1 | 0% | 1,117 | 1,936 | +73% | 0 | 0 | — |
case-16 | fail→pass | 9,026 | 3,032 | -66% | 1 | 1 | 0% | 1,637 | 1,801 | +10% | 0 | 0 | — |
case-17 | fail→pass | 13,281 | 2,604 | -80% | 1 | 1 | 0% | 2,121 | 1,741 | -18% | 0 | 0 | — |
case-18 | pass→pass | 9,395 | 4,424 | -53% | 1 | 1 | 0% | 1,483 | 2,024 | +36% | 0 | 0 | — |
case-19 | fail→pass | 10,098 | 1,440 | -86% | 1 | 1 | 0% | 1,595 | 1,479 | -7% | 0 | 0 | — |
case-20 | fail→fail | 15,245 | 12,099 | -21% | 1 | 1 | 0% | 2,563 | 3,391 | +32% | 0 | 0 | — |
case-21 | fail→pass | 16,270 | 7,476 | -54% | 1 | 1 | 0% | 3,612 | 2,630 | -27% | 0 | 0 | — |
case-22 | fail→fail | 14,050 | 8,778 | -38% | 1 | 1 | 0% | 2,481 | 2,828 | +14% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 19 counted toward the lift figure. The other 3 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +59 percentage points is the difference between those two pass rates over the 19 comparable cases. 1 case got worse with the skill loaded, and it is included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.