Redesign an existing website to make it better. Built on top of impeccable.
88%
Does it follow best practices?
Run evals on this skill
Adds up to 20 points to the overall score
Low
Low-risk findings worth noting
The per-page checks Phase C runs before flipping a page to deployed, and the
batched-delivery flow that runs them uniformly at scale. SKILL.md Phase C names
each gate in one line; the mechanics live here.
Two halves. The gates split into a static pre-PUT lint
(reference/delivery-lint.md → scripts/delivery-lint.mjs: wrapper,
one-CTA-per-<p>, trailing-slash, path-safety — deterministic and offline) and a
dynamic post-deploy check (scripts/verify.mjs § typed render-truth: does it
actually render, typed by page/fragment/index). Run the static lint first — it
catches the cheap failures before a network round-trip. Gate 2 (image-fidelity)
has its own network resolver, scripts/media-reconcile.mjs
(migrate/reference/media-reconciliation.md).
A migration reproduces the source; it must not invent sections. Invented
sections render as empty placeholders or, worse, get back-filled with fabricated
facts. This recurs as trailing cross-link rails (related-*, *-teasers),
specialist/teaser grids, and generic trailing CTAs appended to leaf pages
regardless of source.
node skills/rollout/scripts/section-fidelity.mjs \
--file <content.html> --source <sourceUrl> # authored vs source, side-by-sideThe helper is a scaffold, not a judge: it lists authored block sections
(pre-flagging known filler shapes as ⟵ REVIEW) against the source heading
outline. For each authored section decide:
deployed: the section carries FABRICATED facts
(invented names, made-up events/dates, boilerplate not on the source).
Fabricated content on a real site is the worst migration defect.Delete the section from the content file (and git rm the block if it becomes
orphaned). Genuinely-missing real content is different: leave the block, render
gracefully, log it as a content gap — do NOT invent filler.
<img> src must RESOLVE, or be omitted)The #1 recurring defect at scale: an authored external image URL the preview
ingester can't fetch delivers as <img src="about:error"> — a silent break that
"it renders" hides. The systematic resolver is media-reconcile.mjs (it decides
optimize/keep/rewrite/omit per image and can --apply the fix). Content-host URLs
(content.da.live) are its hosted decision: verified offline against
stardust/deploy/media-ledger.json, never anonymously fetched (a 401 there is by
design); missing from the ledger fails the gate. The manual form, for a single
image, is a 200 check:
node skills/rollout/scripts/media-reconcile.mjs --file <html> --deploy-host <host> # all images
curl -s -o /dev/null -w '%{http_code}' <url> # one imageIf it isn't 200, OMIT the image (the block renders without it) rather than ship
about:error. Never author a logo/placeholder stand-in as if it were editorial.
Two failure signatures where the asset exists behind a malformed URL — fix the
URL, don't drop the image:
…/4x3/768/… 404s; the …/original/768/… sibling
resolves). Rewrite to the resolving variant.…/<id>&wid=600&hei=… (no
?) makes the whole <id>&wid=… a bogus asset id → 403. Repair the first &
after the id to ?.After preview, the authoritative check is .plain.html: 0 about:error and the
expected <img>+alt count (CSS-background images are absent from it).
Real source URLs are not always valid EDS resource paths; DA accepts the PUT
(201) but preview/serve then 404s/400s, so this is invisible until verify. Before
deploy, normalize each path and, when it changes, record the original→normalized
pair so the source URL can be redirected (a final migration must not 404 inbound
links). Rules:
-/_ in any segment;_→- and runs of -;-- segment delimiter (e.g. klinik-st--anna) — AEM reserves --
as the branch--repo--owner host delimiter, so a -- in a path 400s.Append each change to stardust/redirects.tsv (source<TAB>destination); wiring
those into the EDS redirects config is a Phase D/assembly step. When a spec exists
(stardust/spec/knowledge/), run seed-redirects.mjs first: it adds every source
URL that changes on EDS and points the source's own redirects at the delivered
page (deploy/scripts/eds-path.mjs is the path contract), only toward delivered
pages, so re-run it after each wave; rows already in the file win.
At ~1000-page scale the roster comes from the source sitemap, which includes URLs that 404 on the source (stale entries) and pages with no HTML body (PDF-only publication entries — a title plus a PDF download). Two rules:
knowledge/urls.jsonl
says: http-*, redirect-broken, loop dead, flag=empty bodyless; probe blocked
rows). A dead source URL is not a
page to fabricate — leave it un-authored and let it show as the lone gap in the
dashboard (e.g. 814/815). Never invent a body to fill the slot.Separate the concerns so an agent crash or token blowout can't corrupt state, and so the gates run uniformly:
PUT+preview loop):
token stays in one place, retries are trivial, and a sub-agent dying mid-response
leaves its files already on disk. The replica Phase 5 fan-out
(../../replica/reference/handoff-contract.md § 3 row C) is the one sanctioned
variant: each cluster subagent drives its own deploy-batch.mjs --paths … --ledger stardust/deploy/ledger-<cluster>.json --concurrency 2 and gates on the
preview origin itself, which is safe only because the shared ledgers are written
under a lock (coverage, media) or per cluster (deploy), the token is read from
the environment by every driver, and a persistent 401 halts that cluster's batch
(exit 3) instead of failing pages — a recorded 36-page run delivered 36/36 this
way. Outside that contract, author-only + central deploy stays the rule.<h1>, the body/<main>/<footer> wrapper, balanced
<div>s — catches a truncated/garbled file before it reaches DA.deployed after the rendered
.plain.html passes (HTTP 200, 0 about:error, <h1> present). Verify is where
the image- and path-fidelity defects surface at scale — a 4-page sample won't show
them; a 130-page batch will.PUT+preview+live ALL 200 yet 404 at the delivery URL.
The proven case: an uppercase segment (.../CAR-T-Zellen) — the admin API
accepts and "publishes" it, but *.aem.live/.../CAR-T-Zellen 404s (delivery is
lower-cased). The PUT=201 PRE=4xx heuristic MISSES it; Gate 3's "lowercase the
whole path" prevents it at author time, and the GET-.plain.html verify is the
only thing that catches it after the fact. Fix = rename to the lowercase path,
redeploy, DELETE the stale uppercase DA source, add a redirect.PUT+preview loop exceeds a
2-min foreground budget); log per-page OK/FAIL and re-drive only the FAILs.
Transient PUT=000 → retry; PUT=201 PRE=4xx/400 → a path-safety case (Gate 3);
200 + about:error → an image case (Gate 2).node/curl inside a multi-line while/for can lose PATH
("command not found") — write the loop to a bash script file with absolute
binaries and run it, rather than inlining..tessl-plugin
skills
audit
reference
deploy
reference
scripts
test
diff
direct
dynamics
extract
migrate
prepare-migration
prototype
assets
motion
fixtures
reference
qa
replica
reference
scripts
test
reskin
rollout
spec
stardust
uplift
reference