DeepSeek Harness Remote
DeepSeek Harness Remote 可将手机或另一台电脑与你运行 DeepSeek Harness 的电脑安全连接,随时随地继续你的工作,方便随时查看进度并继续工作。
Translation pending
Describe your task and copy a complete prompt that asks the AI for consistent terms, evidence kept apart from inference, and questions instead of guesses.
ProductivityGitHub @HeliakonOnlineOwner unconfirmed
Have you tried it? One tap is enough. No sign-in needed.
Repository information has not been collected yet.
No updates detected yet. GitHub releases will appear here.
Write prompts for clearer AI answers. Describe your task, choose the shape of the answer, and copy a complete prompt. Your entered text is processed locally and is not uploaded. Also: editing briefs for natural writing and coding rules for Claude Code.
Taken from the repository README
Loading the full README.
Read the full README ↗A static, five-page site for writing better AI prompts:
/): turns a task description, task type, answer format, and optional source material into a complete prompt./humanize/): turns readers, original text, and an optional voice sample into an editing brief in English, Polish, or Portuguese./claude/): a concise, coding-focused CLAUDE.md to preview, copy, or download./claude-outputs/): a short guide to Claude Code output styles, and when CLAUDE.md fits better./evidence/): sourced checks of the claims behind the site, plus a dated review of eight skill and plugin repositories.The site composes prompts in the browser. It does not call an AI model, and no visitor text is sent over the network.
Every release also has a krakem.html.sha256 file. To check your download, run one of these and compare the result with it:
Get-FileHash .\krakem.html -Algorithm SHA256
shasum -a 256 krakem.html
content/: page bodies. Edit these, not dist/.assets/: styles, scripts, artwork, the social preview image, and downloads/CLAUDE.md.
prompt.js and humanize-brief.js compose prompts. guide.js and humanize.js connect them to the forms.drafts.js keeps unfinished form drafts in sessionStorage for the current tab.field.js and field-shapes.js draw the particle field.common.js runs the menu, the Motion toggle, and copy buttons. screensaver.js shows the idle screensaver.site.config.json: the public site URL, used for canonical URLs and social preview tags..github/workflows/pages.yml: builds, checks, and deploys to GitHub Pages..github/workflows/release.yml: when a release is published, attaches krakem.html and its SHA-256 checksum.scripts/build.mjs: assembles dist/.scripts/build-offline.mjs: turns dist/ into release/krakem.html, one file with every page, style, script, and download inline.scripts/check.mjs and scripts/check-html.py: automated checks.scripts/brand-mark.mjs: the particle asterisk mark. The build inlines it in the header and footer; run it directly to regenerate assets/icon.svg and the mark in scripts/social-preview.html.scripts/social-preview.html: source for assets/social-preview.jpg.evaluation/: saved prompts, answers, and grades from two small evaluations.node scripts/build.mjs
node scripts/build-offline.mjs
node scripts/check.mjs
python3 scripts/check-html.py
python3 -m http.server 8765 --directory dist
Serve dist/ over HTTP. Opening index.html directly from disk does not work: browsers block module scripts on file:// pages, and folder links do not resolve to index.html.
The site is set up for GitHub Pages at https://heliakon.github.io/krakem/.
.github/workflows/pages.yml runs on every push to main. It builds dist/, runs both check scripts, and deploys only if they pass. Pull requests are built and checked but not deployed.krakem for the URL above.dist/ is built in CI and ignored by Git. Edit content/ and assets/, never dist/./humanize/). The build turns them into relative links, so the same output works at a domain root or under /krakem/. check-html.py fails on any root-relative link or CSS url() left in dist/.siteUrl in site.config.json sets canonical URLs and social preview tags. If the site moves, change it there, or set SITE_URL for one build.After the first deploy, open https://heliakon.github.io/krakem/assets/social-preview.jpg and https://heliakon.github.io/krakem/assets/downloads/CLAUDE.md in a private window to confirm anonymous access.
To regenerate the 1200×630 preview image after a design change:
"/Applications/Google Chrome.app/Contents/MacOS/Google Chrome" --headless=new --hide-scrollbars --force-device-scale-factor=1 --window-size=1200,630 --virtual-time-budget=2000 --screenshot="$PWD/social-preview.png" "file://$PWD/scripts/social-preview.html"
sips -s format jpeg -s formatOptions 86 social-preview.png --out assets/social-preview.jpg
Prompt creator. The creator is the first section below the menu. Prompts include the visitor's task, task-specific guidance, an answer format, and any source material, with no editing placeholders. Source material is included exactly as entered. Comparison mode requires two different subjects, a decision goal, and criteria before copying. Missing-detail links focus the relevant field. Summaries and checklists are available alongside the existing task and format options. Choosing a task sets a suitable answer format until the visitor picks a format explicitly.
Humanize. Readers and original text are required before copying. The voice sample is optional; when it is empty, its section is left out. The original text is inserted exactly as entered, as the last section of the brief.
Drafts. Creator and Humanize inputs are saved to sessionStorage as the visitor types, including task type, format, comparison fields, and the explicit-format flag. They survive navigation and reload in the same tab and are discarded when the tab closes. Clear draft resets the form and removes the stored draft. If storage is blocked, the forms still work and the page says drafts are not kept. localStorage holds only the motion preference.
Particle field. A full-width field follows the creator: 780px tall on desktop, 620px on tablets, and 560px on phones. Seven unnamed shapes morph every five seconds, starting with a nautilus and a crab. A close camera crops the forms on purpose. Motion and Chaos controls sit beside the animation, and the numbered buttons work with motion off. Rendering respects reduced-motion preferences. It is capped at DPR 1.5 and pauses when offscreen or in a hidden tab.
Menu. The header menu is a dock of five buttons. Each shows an icon, a name, and a short description. With a mouse or trackpad and motion on, buttons grow as the pointer approaches. On phones, the descriptions are hidden and the dock spans the full width.
Screensaver. After 12 seconds without input (pointer movement, clicks, keys, scrolling, or touch), a Matrix-style rain fades in over 0.8 seconds. Then the page underneath is hidden so it stops painting, and the particle field pauses. Three small controls, placed as in the original screensaver file, set letter speed (top left), fade-in mode (the checkbox), and color (top right, from the site green through blue, purple, orange, and red to white). Moving the pointer keeps the rain so the controls can be reached. A click outside the controls, a key, Escape, or the scroll wheel brings the page back. The click or key that does this does not act on the page, and keyboard focus returns to where it was. It does not start on touch screens, with reduced motion or Motion off, while a form field has focus, or in a hidden tab. The rain is drawn at one canvas pixel per CSS pixel. The standard mode draws a step every 10 to 120 ms, set by the speed control; fade-in mode redraws at most about 30 times a second.
Evidence. The page keeps its original research and review dates. GitHub star counts are a 23 October 2026 snapshot of popularity, not a quality or security signal. All eight reviewed repositories remain listed as historical research. Humanize, Claude.md, and Claude outputs do not recommend or install skills.
CLAUDE.md. Edit assets/downloads/CLAUDE.md. The build inserts that file into the Claude.md page preview and copies it to the download path. It has no external imports.
scripts/check.mjs covers:
These are regression checks. They do not establish full accessibility conformance, real-device performance, or model-output quality. Small model evaluations are recorded in evaluation/; they are not benchmarks.
MIT. See LICENSE.