CtrlK
BlogDocsLog inGet started
Tessl Logo
PluginPublic

adobe/stardust

Redesign an existing website to make it better. Built on top of impeccable.

Quality

88%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide
SecuritybySnyk

Low

Low-risk findings worth noting

Overview
Quality
Evals
Security
Files

reading-discipline.mdskills/replica/reference/

Reading and inspection discipline — the context is the budget

Read from the replica card (../SKILL.md § Setup) when a step is about to read a file, a capture or an instrument's output.

Everything a step prints stays in the agent's context for the rest of the session and is re-read on every model call. Recorded in one hands-off session (2026-09-18, 77 minutes): 172 steps carried 750k characters of tool output into a 535k-token context; three whole reference documents cated at once overflowed and were then read again in full (157k characters for text paid for twice); 21 ad-hoc node -e dumps of capture JSON cost 141k, sed -n line ranges of a pretty-printed stylesheet 102k. The helpers below print the part that matters, capped, and say what they left out:

  • Reference docs: node stardust/scripts/replica/section.mjs <doc> --list for the outline, then section.mjs <doc> '<heading>' for the one section the current step needs — never the whole file. One section per call: a recorded call for five sections returned 82 KB that stayed in context for 160 turns (--all over several sections is capped at 20 KB now).
  • Captured CSS: css-rules.mjs <sheet> '<selector regex>' [--media <re>] [--decl <re>] — the rule blocks, each with its media condition, instead of line ranges of the sheet.
  • Capture JSON (computed styles, content trees, motion checks, crawl logs — their shape is the run's own): json-query.mjs <file> for the shape, --path <p> --keys to learn an array's fields, then --path <p> --match <key>=<re> --fields <a,b> --max 40 for a bounded table; values that feed a command (URLs, slugs, selectors) come from --tsv or --path <record>.<field>, never from the table — its cells are capped.
  • Captured HTML: html-slice.mjs <page.html> header|footer|main|.class [--text] — one element, attributes stripped to the structural few, scripts and inline SVG removed, capped.
  • Instrument output: never cat a log or a capture; gate rounds run through run-bg.mjs (Phase 4) and are read back with wait summaries and log --grep.
  • Live vs prototype boxes: measure.mjs <live-url> --against <prototype-url> --selectors "<css>,…" [--width 1440,360] — rect + computed values per selector, one delta line per box (Phase 3); never an authored probe or a one-off node -e evaluate.
  • Images: one crop per fact, never a full stitched page (it is downscaled past legibility and costs a step), never the live/build/diff triplet — the diff crop of the first hot band (crop-compare.mjs) is the one image a round needs. Recorded: 21 image reads in one run, most of them full pages or triplets, for facts the verdict lines already held. Step 10's eyeball of the DEPLOYED page is the published gate.sh --full round plus header/footer crop-compare.mjs bands (the two-band invocation and where each band number comes from: gate doc § Pass bar, item 5), not a full-page image. The one exception is the brand-gestalt read at extract, which uses the shipped thumb.mjs (skills/extract/scripts/thumb.mjs, box-filtered to 480 px or narrower under its size cap; a line with a crop share below 100 is followed by its --offset slice) — the whole page per archetype, never the raw capture.
  • Script flags: every shipped CLI script — replica, diff, master, deploy, rollout, dynamics, qa, extract, reskin — answers --help (and -h) with its usage header before it parses anything, touches a file or opens a browser; the header names the flags, the exit codes and, for a script that writes artifacts, a Writes: block listing every path (the script-help lint keeps this true). The Phase 5 contract card's usage table (reference/handoff-contract.md § 4) remains the one place with the full invocation shape per deploy and rollout script. Never sed/head/ grep a script's source to learn its flags or what it writes (16 such reads in one recorded session, four greps over one crawler in another): run it with --help. Index first: ../stardust/reference/scripts-index.md has every shipped script on one line (what, key flags) — ask --help only for a flag it does not name (a recorded run asked --help 78 times).

Write anything you will need again to a file under stardust/ (a lifted value table, a section map) and read the file back by query, not the instrument's output by scroll.

skills

README.md

tile.json