Communitygithub.com

youtube-thumbnail:生成、評分、修改 YouTube 縮圖

CoverHook 推出的 Agent 技能:依影片主題與可選照片設計高點擊的 YouTube 縮圖,按頻道類型規則給 100 分制分數,最多自動修改兩次。呼叫 CoverHook MCP 服務;生成按點數計費,失敗自動退款。

youtube-thumbnail:生成、評分、修改 YouTube 縮圖 是什麼?

YouTube 單集配方第 2 步(縮圖),來自 coverhook-ai/youtube-thumbnail-skill(2026-10-09 發布,技能檔案 MIT;CoverHook 服務按張收費,新帳號送 25 點免費額度)。給它影片主題、縮圖上要放的三到五個字詞,最好再加一張人臉或產品照片。Agent 先呼叫 plan_cover 建議類型並精簡標題,再用帶冪等鍵的 generate_cover 生成(重試不會重複扣點),每 10–15 秒查詢一次 get_cover,接著 score_cover 給出 100 分制分數與未通過的檢查項。低於 70 分就以一兩條具體意見(如「臉放大」「字少一點」)呼叫 refine_cover,未經詢問最多改兩次,改了分數沒漲就停。預設每次只出一張,要備選才多出。同一外掛還附一個免帳號的免費縮圖點評,以及頻道橫幅與書封技能。

相容平台~Claude Code~Codex CLI~Cursor
npx skills add https://github.com/coverhook-ai/youtube-thumbnail-skill/tree/HEAD/plugins/coverhook/skills/youtube-thumbnail

在你喜歡的 AI 中提問

開啟一個已預先載入此 Agent Skill 的新對話。

說明文件

YouTube thumbnail

CoverHook designs the thumbnail; your job is to get the right inputs, run the tools in order, and show the results. The design rules for each channel niche live on CoverHook's side and are applied when the image is generated — you do not write image prompts.

What to find out first

Ask only for what is missing, in one message:

  1. The video. Its topic, and the few words that should appear on the thumbnail. Keep on-image text short: three to five words reads at feed size, a sentence does not.
  2. The channel's niche. Call list_skills to see the niches and which ones work best with a photo. If the person has not said, plan_cover will suggest one from their description.
  3. A photo, if they have one. A face or a product shot usually makes a stronger thumbnail. If they give you a local file, send it with upload_image and use the returned URL.

If the request is clear enough, skip the questions and start.

Workflow

  1. plan_cover with the person's own words. It returns a suggested niche, a tidied on-image title, and any questions still worth asking. Ask those questions only if the answer would change the design.
  2. generate_cover with the niche, title, and any reference image URLs. It returns a task id straight away and charges credits; the image takes about a minute. Always pass an idempotency_key so a retry never charges twice.
  3. get_cover until the task is completed or failed. Poll every 10 to 15 seconds, not faster.
  4. score_cover on the finished task. It returns a score out of 100 and the checks the thumbnail failed.
  5. If the score is below 70, or the person asks for changes, refine_cover with one or two concrete notes drawn from the failed checks or their feedback ("make the face larger", "fewer words"). Refine at most twice without asking; if the second revision does not improve the score, stop and let the person choose. A score of 70 or more is done: mention any failed check, but do not refine on your own — every revision costs credits.

Make one image per request. Only when the person asks for options, call generate_cover two or three times with different strategy values from list_skills. The account runs at most three generations at once.

When a generation fails

A failed task has already been refunded; say so, with its error in plain words.

  • It took longer than 3 minutes, or the service was unavailable: offer to try again once. Use a new idempotency_key — the old one returns the same failed task.
  • The request was declined: do not retry the same thing. Suggest a different title or photo.
  • A reference image could not be read: ask for the image again and upload it with upload_image.

Never retry more than once without asking.

Showing results

Show the image, its score, and one line on what to change if anything. Do not paste the tool output. Mention what the run cost when it is more than one generation; get_credits shows the balance.

Principles that hold across niches

  • One clear subject. A busy thumbnail loses at feed size.
  • High contrast and saturated colour, so it stands out next to the thumbnails around it.
  • Few, large words that do not repeat the video title.
  • Make the viewer curious about the video, and stay honest about what is in it.

If the CoverHook tools are missing

This skill needs the CoverHook connection: tools such as list_skills and generate_cover. If they are not available, or the CoverHook server failed to connect (a 401, "not authenticated"), the usual reason is that no API key has been entered yet: the plugin installs without one. Do not assume a key was revoked or mistyped, and do not work around it by writing an image prompt yourself. Tell the person, in one short message that mentions the free credits:

  1. Sign in at https://coverhook.com/account and create an API key in the API keys section. New accounts get 25 free credits, enough for one thumbnail at the default size.
  2. Run /plugin, open CoverHook on the Installed tab, choose Configure options, and paste the key.
  3. Run /mcp to reconnect, or restart the session.

If they say they already entered a key, it may have been revoked or pasted with a typo: create a new one and replace it the same way.

If they connected CoverHook as a plain MCP server instead of this plugin, the key goes in the Authorization: Bearer header of that server.

Individual skills in this repo

This repo contains 3 individual skills — each has its own dedicated page.

相關技能