adobe/snowflake

Snowflake — Static-to-EDS Conversion

Convert a static HTML page into an EDS page while preserving the original design and making content authorable in Document Authoring. Two conversion levels are supported:

• Page-level (overlay) — the original DOM is preserved byte-for-byte via a template with [data-slot] markers. One template, one CSS file, one DA doc with slot-keyed rows.
• Block-level — each content section becomes an independent EDS block with its own decorate() function and CSS. Content is authored in DA block tables. Header and footer stay as static fragments.

Both levels keep the original visual design intact. Page-level is the safer default; block-level produces more standard EDS output but requires section independence in the source page.

When to use

The user has an AI-generated polished static HTML page and wants to launch it on Edge Delivery Services without losing the original design while still making content editable in DA. Typical phrasing:

• "Convert https://example.com/static-page to EDS"
• "Make this page editable in DA but keep the original markup"
• "Convert this page to EDS blocks" (signals level=block)
• "Start the next experimentation for URL …"
• "Static-to-EDS overlay for …" (signals level=page)

What this skill does NOT do

Not for canonical EDS block-rewrite migrations — that's page-import. Snowflake preserves the source design; page-import rewrites to standard EDS block patterns with no visual fidelity target. Three asset strategies are supported (see knowledge/methodology.md §3): absolute, vendor, da-media.

Parameters

| Parameter | Values | Default | Description | |-----------|--------|---------|-------------| | level | page, auto, check, block | page | Conversion level — see below |

level values

| Value | Behavior | |-------|----------| | auto | Run the feasibility analysis in Phase 2, present the recommendation, ask the user to confirm before Phase 3 | | check | Run the feasibility analysis only — produce the report in decisions.json, stop before Phase 3. Useful for batch scanning | | block | Block-level conversion. Analysis still runs as validation (written to decisions.json) but does not gate the conversion | | page | Page-level conversion (standard overlay). Analysis still runs as validation but does not gate the conversion |

When level is not provided and the user's phrasing signals intent, infer it:

• "convert to EDS blocks", "block-level" → level=block
• "overlay", "preserve the DOM", "snowflake overlay" → level=page
• Neutral phrasing → level=page

Usage examples

/snowflake https://example.com/promo              → page-level; infer repo, daRoot, slug
/snowflake https://example.com/promo level=block  → block-level, infer the rest
/snowflake https://example.com/promo level=auto   → feasibility analysis decides
/snowflake level=check                            → feasibility scan only (asks for URL)
/snowflake                                        → page-level, fully interactive

The Source URL is the leading positional input and the only required argument. Everything else is resolved automatically and presented in a single confirmation summary before any work begins.

Skill dependencies

Snowflake cites DA HTML rules and the DA admin API contract from the da-content skill. Load da-content alongside Snowflake. Phases 3 (Generate) and 5 (Round-trip) reference it directly.

Prerequisites

Required — the only input the skill cannot resolve on its own:

1. Source URL — the static page to convert. Must be reachable (publicly hosted or local dev server).

Resolved automatically — shown in the init summary for one-shot confirmation before any work begins:

2. Target EDS repo — detected via gh repo view --json nameWithOwner, falling back to parsing git remote get-url origin. Must already have the overlay engine wired (see knowledge/architecture.md §"Solution shape"). Phase 0 installs it if absent.
3. DA root path — read from .snowflake/config.json daRoot key if set, otherwise defaults to the current git branch name (the same branch the skill uses for code). Shown in summary; override inline.
4. Conversion level — inferred from phrasing (see Parameters), else page. Shown in summary; override inline.
5. Slug / template name — derived from the source URL (kebab-case, ≤30 chars). Shown in summary; override inline.

Auth check (non-blocking) — DA token resolved from $DA_TOKEN → ~/.aem/da-token.json. Its status appears in the init summary. Phases 1–4 do not need it; if absent at invocation time, invoke the da-auth skill before Phase 5 (Round-trip) runs.

Initialization

On every invocation the agent performs these steps before entering Phase 0:

1. Resolve inputs — apply the fast-path rules from the Prerequisites section above.

2. Probe substrate —

   node <SKILL_DIR>/scripts/install-substrate.mjs --dry-run
   

   Captures the outcome (no-op / clean-install / drift / custom-code-detected).

3. Check DA token —

   DA_TOKEN=$(node -e "
     const fs = require('fs');
     const p = process.env.HOME + '/.aem/da-token.json';
     try {
       const t = JSON.parse(fs.readFileSync(p, 'utf8'));
       if (t.expires_at > Date.now() + 60000) process.stdout.write(t.access_token);
     } catch {}
   ")
   

   Non-blocking — records status for the summary only.

4. Always display the run parameters before any phase begins. Show this summary unconditionally — even when all values were provided upfront or inferred without ambiguity — then proceed immediately without waiting for confirmation:

   Source URL : https://example.com/promo   ← required (provided)
   Target repo: acme/my-site                ← detected from git
   DA root    : /main                        ← from current branch
   Level      : page                        ← default
   Slug       : promo                       ← derived from URL
   Substrate  : clean install — 9 files     ← (or: already current ✓)
   DA token   : cached ✓                    ← (or: not found — needed at Phase 5)
   

   Proceed without pausing for the common case (fresh install — no snowflake substrate yet; replaced files are backed up). Only pause for drift (a prior snowflake substrate that diverged), where overwriting could lose intentional customization — Phase 0 handles that case.

5. Proceed — Phase 0 → 1 → … → 6 in order.

Quick start — end-to-end example

From the target EDS repository root, here's the full seven-phase conversion in compressed form. Each phase file under phases/ holds the complete prompt; this is the shape of the actual commands the agent emits.

# Inputs (resolved during init, confirmed in the summary)
SOURCE_URL="https://example.com/promo"
PAGE_SLUG="promo"
DA_ROOT="/main"                       # defaults to current branch name
NNN=001                           # next run number
PROJECT=".snowflake/projects/${NNN}-${PAGE_SLUG}"
TEMPLATE_NAME="promo"
LEVEL="page"                      # page | auto | check | block

# Phase 0 — install (or verify) the overlay substrate (once per repo)
node "<SKILL_DIR>/scripts/install-substrate.mjs"

# Phase 1 — capture: fetch source + assets into the project folder
mkdir -p "$PROJECT/input"
curl -fsS "$SOURCE_URL" -o "$PROJECT/input/index.html"
# For JS-rendered pages, use a browser to get the fully rendered HTML instead

# Phase 2 — analyze: produce decisions.json (sections, slots, asset
# strategy, head-links, conversionLevel). Includes block-level
# feasibility assessment. If LEVEL=check, stop here.
# Driven by phases/2-analyze.md.

# Phase 3 — generate: produce 5 artifacts + DA-source body
# (templates/<tpl>.html, fragments/<tpl>/{header,footer}.html,
# styles/<tpl>.css, scripts/<tpl>-animations.js, da/<slug>.html).
# Driven by phases/3-generate.md.

# Phase 4 — wire: copy artifacts to EDS-served paths and build the
# drafts file. Driven by phases/4-wire.md.

# Phase 5 — round-trip: local dev server + production preview
npx -y @adobe/aem-cli up --html-folder drafts &
TOKEN="${DA_TOKEN:-$(jq -r .access_token ~/.aem/da-token.json)}"
git checkout -b "snowflake-${NNN}" && git add . && git commit -m "snowflake #${NNN}"
git push -u origin "snowflake-${NNN}"
curl -X PUT -H "Authorization: Bearer $TOKEN" \
  -F "data=@${PROJECT}/output/da/${PAGE_SLUG}.html;type=text/html" \
  "https://admin.da.live/source/${OWNER}/${REPO}${DA_ROOT}/${PAGE_SLUG}.html"
curl -X POST -H "Authorization: Bearer $TOKEN" \
  "https://admin.hlx.page/preview/${OWNER}/${REPO}/snowflake-${NNN}${DA_ROOT}/${PAGE_SLUG}"
# Verify at https://snowflake-${NNN}--${REPO}--${OWNER}.aem.page/${DA_ROOT}/${PAGE_SLUG}

# Phase 6 — reflect: append findings to $PROJECT/learnings.md;
# promote cross-project rules to knowledge/learnings.md.

<SKILL_DIR> is the absolute path to the directory containing this SKILL.md. The agent substitutes it before invoking — see HOST-NOTES.md for per-host resolution rules.

The seven phases (sequential)

Each phase is a self-contained markdown file with executable bash + Node. The agent reads the phase prompt, runs its steps, updates state.json at the project root (<projectsDir>/<NNN>-<slug>/state.json), and proceeds. Reruns are safe — phases skip work already done.

0. Prerequisites — install/verify the overlay substrate; stamp .snowflake/config.json. Runs once per repo. See phases/0-prereq.md.

1. Capture — fetch source HTML and referenced external assets; create the project folder. See phases/1-capture.md.

2. Analyze — structural map: header/footer boundaries, section list, slot opportunities, head-level links to lift, asset strategy. Also runs the block-level feasibility assessment — see knowledge/block-level-feasibility.md. Includes a page complexity gate: pages with >8 sections or >100 slottable elements auto-switch from page-level to block-level to avoid incomplete content extraction (see phases/2-analyze.md). Produces notes.md + decisions.json (including conversionLevel). See phases/2-analyze.md.

3. Generate — branches by conversionLevel from Phase 2:

   • page: produce the 5 overlay artifacts (template HTML, header fragment, footer fragment, page CSS, animations JS) plus the DA-source body with slot-keyed rows.
   • block: produce per-section block JS/CSS, header/footer fragments, global styles/tokens, and the DA-source body with standard block tables.
   • hybrid: block-level for passing sections, page-level fragments for failing sections. Both page and block are supported under either substrate flavor (eds or milo). All Milo-specific deltas — chrome metadata, page-level/block-level generation, the --pa-* animation sidecars, and wiring — live in assets/substrate-milo/FLAVOR.md; the phase docs carry a gated pointer to it. The core skill stays substrate-neutral. See phases/3-generate.md.

4. Wire — copy artifacts to EDS-served paths, build the local-test drafts file, run lint. See phases/4-wire.md.

5. Round-trip — local (dev server) then production (branch + push + DA PUT + preview API). Enforces a browser health gate on both: the page must render (not blank), apply the overlay, match the source structure, be free of console/network errors, and pass the 1:1 DOM-equality check before the run may continue. See phases/5-roundtrip.md.

6. Reflect — append run findings; promote cross-project learnings to knowledge/learnings.md. Does not close the iteration — that's a user decision. See phases/6-reflect.md.

Knowledge resolution per phase: each phase tries .snowflake/knowledge/<file>.md (project-specific override) first, then <SKILL_DIR>/knowledge/<file>.md (bundled, canonical). Project overrides win on conflict.

Reading order

Load knowledge just-in-time — only when the phase that needs it begins.

On invocation (before Phase 0):

1. This file.
2. knowledge/methodology.md — canonical phase rules.
3. Confirm the da-content skill is loadable (needed by phases 3 and 5).

At Phase 2 (Analyze): 4. knowledge/block-level-feasibility.md — criteria for block-level vs page-level conversion.

At Phase 3 (Generate): 5. knowledge/architecture.md — overlay engine and slot writer semantics. 6. knowledge/block-level-conversion.md — architecture and patterns for block-level conversion. 7. knowledge/learnings.md — cross-project findings relevant to generation.

At Phase 5 (Round-trip): 8. knowledge/eds-da-mechanics.md — EDS pipeline overlay-runtime lore. 9. knowledge/learnings.md — entries on media handling, CORS, scroll-animation quirks (if not already loaded).

Then start at Phase 0.

Further reading (not loaded by the agent)

• README.md — human-readable overview, install commands, contribution guidelines.
• HOST-NOTES.md — per-host adapter notes (Slicc, Claude Code, generic shell), <SKILL_DIR> path resolution rules, .snowflake/ directory convention, and forbidden cross-host primitives (for maintainers).
• examples/README.md — pointers to worked examples from closed iterations.