本文へ移動
cccskills
無料GitHub で公開

canvas-house-edit

Edit a house design instead of rewriting one: the twelve proven designs in canvas_read's gallery are all perfect, so start from one, keep its composition and its look, and change only what the person asked for - proving each round with canvas_audit and a look at the render. Use when the person asks to change, tweak, adjust, restyle or fix a banner that already exists, or asks for a new banner of a kind the gallery already carries.

インストール方法を見る

含まれるファイル(3)

  • SKILL.md5.7 KB
  • reference/copy.md7.8 KB
  • reference/gate.md3.5 KB

SKILL.md(原文)

インストールする前に、エージェントに与えられる指示の中身を確認できます。

Editing a house design

The gallery is the answer to "start from nothing". canvas_read lists twelve house examples, each a destination preset, a proven composition and a look that have already been chosen well - and every one of them passes the gate in reference/gate.md: no lint of any kind, a 2:1 type ratio, a single focal point, and the words inside the budget its own typography allows. That is what "perfect" means here, and it is measured rather than asserted.

So the first question is never "what should I draw". It is:

canvas_read                                  # the gallery, the archetypes, the styles
canvas_new { example: "night-launch" }       # ONE of them, opened as a design you own
canvas_audit { id: "<the new id>" }          # the gate, with no page open
canvas_render { id: "<the new id>" }         # the picture, and the 25% feed thumbnail
read_image <the PNG the report names>

Edit, do not rewrite

A rewrite throws away the one thing you cannot recover, which is the round trip. So:

  • canvas_patch is the edit. One to sixty-four pointer operations on the stored document, validated exactly like a write. Use it for every change: a word, a colour, a size, a position, a reorder, a removal.
  • canvas_write is for a document you actually intend to replace. If you find yourself re-emitting the whole design to move one element, you have taken the long road and lost the composition's own alignment in the process.
  • Patch ONE thing, then look. Ten small patches with a render between them beat one large patch nobody looked at. The gate and the render report are cheap; being wrong about eight changes at once is not.
  • The composition is not yours to improve. The house designs were composed by somebody who had the numbers in front of them. Change what was asked for; if you believe the layout itself is wrong, say so and render the evidence.
  • A style is not a composition. canvas_style re-colours and re-types and never moves anything - so "make it feel calmer" is a style or a palette patch, and "make the headline bigger" is a patch on tokens.scale.display, not a restyle.

The gate, and how to use it

canvas_audit runs five checks and names the ones that fail. Nothing is finished until it passes all five, and it costs nothing: it lays the design out on the host with the same engine the browser uses, so no page has to be open.

CheckWhat it measuresWhat to do when it fails
SYNTAXthe document validatesfix the named field; the validator's code names the path
LINTeverything lintLayout foundthe code names it: MARGIN, SAFE_AREA, LOW_CONTRAST, TYPE_TOO_SMALL, TEXT_OVERFLOW, TEXT_TRUNCATED, TEXT_UNWRAPPED, OFFGRID, SIBLING_EDGE, TEXT_ON_IMAGE, MISSING_ASSET
TYPE_RATIOthe headline against the next sizeraise tokens.scale.display or lower the next size: 2:1 is the floor
COPY_BUDGETthe words against the design's own typecut words, or move them to a subhead - never shrink the type to fit them
FOCALhow many elements share the largest sizemake one of them smaller, or demote it to the caption scale

The gate's numbers are the package's own average glyph, so a BORDERLINE TEXT_OVERFLOW is worth confirming with canvas_render, which measures with the browser's real faces. Everything else - margins, contrast, ratio, the focal point - measures the same either way.

The review protocol

When you have to say whether a design is finished - yours or one you were handed:

  1. The gate first. It is objective, instant and complete. Do not spend a render on a design the gate has already failed.
  2. Then the picture, at both sizes. canvas_render hands back the full PNG and a 25%-scale feed thumbnail. Read BOTH with read_image. A headline that reads beautifully at 1280px can be a grey smear at 320px.
  3. Say what you see, in this shape: the observation (what is actually there), the problem (why it matters at the size it will be seen), and the fix (a specific operation). "The subhead is 22px" is not a finding; "the subhead is 22px, which is 9px in the feed thumbnail, so raise it to 26 or cut it to four words" is.
  4. Name the strongest and the weakest thing. A design has one focal point; if you cannot say which element it is, the design does not have one.
  5. A tie is a defect. Two elements at the same largest size, two buttons, two accents: the gate reports the tie and the eye reports it as noise.

Where the long answers live

ReadWhen
reference/gate.mdevery check the gate runs, what it measures and why - generated from the gate itself
reference/copy.mdthe per-destination character and word budgets, and where each file is uploaded
canvas-design skill, reference/document.mdevery key of every node kind, the token block, the defaults, and every validator code with its fix
canvas-design skill, reference/recipes.mda copy-paste JSON fragment for the surfaces a patch reaches for

レビュー

まだレビューはありません。使ってみた感想をお寄せください。

同じリポジトリのスキル

概要と使いどころ

Design a banner, social preview or poster from nothing: pick the destination preset, then the composition, then write the words, then apply ONE look - and prove it with canvas_audit before you render. Use when the person asks for a banner, an OG card, a social image, a README header, a poster or a launch card and no design exists yet.

日本語の概要は準備中です。原文の説明を表示しています。

vecnode/vncode82026年10月9日 更新

Design a canvas document the Canvas tab and the host both understand: the JSON language, the layout engine's sizing rules, the render report's codes, composition craft with numbers, copy budgets, and copy-paste recipes. A bundled reference carries every field and every validator code.

日本語の概要は準備中です。原文の説明を表示しています。

vecnode/vncode82026年10月9日 更新

Drive ffmpeg through the media_run tool as an argv array - remux instead of re-encode, map streams explicitly, trim at keyframes, scale and crop, extract audio and frames, and build GIFs, contact sheets and concatenations. Covers the flags this pack injects, reading a failure from its exit code and last log lines, and the version-dependent spellings to hedge.

日本語の概要は準備中です。原文の説明を表示しています。

vecnode/vncode82026年10月9日 更新

Read a media file's real structure with ffprobe through media_run - the JSON invocation and how to read it, what every family of field means, how to count frames, sample without decoding everything, detect HDR and interlacing, and diagnose a missing duration, a stream with no frames or a container that lies. Covers when media_probe's own report is already enough.

日本語の概要は準備中です。原文の説明を表示しています。

vecnode/vncode82026年10月9日 更新

Author Mermaid diagrams that parse the first time: choose the right diagram type for the question, write syntax the parser accepts, and keep the picture readable in a narrow panel. A bundled reference covers complex, large diagrams.

日本語の概要は準備中です。原文の説明を表示しています。

vecnode/vncode82026年10月9日 更新

Read and understand PDF documents with the pdf_* tools - inspect a document before reading it, extract text in reading order or as reconstructed layout, search a document without reading all of it, recognize scanned pages that have no text layer, and render pages as pictures. Covers scanned documents, chat attachments, page-range and character caps, and the read-only and untrusted-content rules.

日本語の概要は準備中です。原文の説明を表示しています。

vecnode/vncode82026年10月9日 更新

vecnode のスキルをすべて見る

このスキルの問題を報告する