Generate and edit Draw.io, Mermaid, and Excalidraw diagrams from natural language using a structured JSON spec.
Coding
sync-config-with-sample
Try itSync a production config file (e.g. hyprland.lua.gpd) against the latest upstream sample (e.g. hyprland.lua.260917) by copying only reference info/URL comments while preserving the user's actual configuration. Use when the user drops a freshly downloaded sample alongside a live config, mentions keeping the newest sample as the diff baseline, or wants only real config differences to remain visible in diff.
What it does
The user keeps a **production config** (live file, archived in this repo with a device suffix like ) next to the **latest upstream sample** (freshly re-downloaded from the project's GitHub repo, named with a date, e.g. ). The sample is the baseline: the only acceptable diff is the user's real conf…
The skill document
Sync config with upstream sample
The user keeps a production config (live ~/.config/... file, archived in this repo with a device suffix like .gpd) next to the latest upstream sample (freshly re-downloaded from the project's GitHub repo, named with a date, e.g. .260917). The sample is the baseline: the only acceptable diff is the user's real configuration — never reference noise, never lost config.
Rules (immutable)
- Never modify the sample file. It will be re-downloaded as the example. If it must change, copy a fresh one — never edit it in place.
- Never change actual configuration in the production file: values, binds, active blocks, added features. Copy only reference information.
- Reference info = anything inert: comment lines containing wiki/repo URLs, section headers,
-- See https://...links, wording of explanatory comments that upstream rewrote. These are safe to adopt verbatim from the sample. - The end state is verified by
diff: the only remaining hunks must be real config differences (user's values vs sample's), with zero reference/URL noise. - Raise hands when a hunk is ambiguous (see Ground truth below) before touching anything ambiguous.
Procedure
1. Establish files and intent
- Identify the production file (
*.gpd= this device's live config) and the sample (dated*.2609xx). - State the two "don't touch" sets explicitly up front: the sample, and the user's config.
2. Diff and classify every hunk
diff hyprland.lua. hyprland.lua.
For each hunk, classify into one of three buckets:
- Reference → apply (copy the sample's exact wording into the production file).
- Config → leave untouched, but list it so the user sees what will remain in the diff.
- Ambiguous → pause and investigate (below).
3. Ground truth before acting on ambiguous hunks
A diffed line may look like config but really be an upstream rename, deprecation, or fixed typo. Determine the truth before deciding:
- Read the sample itself (does it document this part?).
- Check upstream: GitHub PRs/commits/issues for the change (e.g.
hyprwm/Hyprland), the wiki page behind the reference URL. - Check what the installed version on this machine understands:
hyprctl version,pacman -Q. If the rename landed after the installed release, keep the old name — it works now and the compat shim usually accepts it after upgrade too. - Present findings + recommendation, get the user's decision, apply.
4. Apply reference edits
Use edit with exact unique oldString matches. Take wording verbatim from the sample (whitespace included). Only touch reference lines — if an edit's oldString overlaps config code, split the edit or re-classify.
5. Verify with the user, step by step
- Re-run
diff hyprland.lua. hyprland.lua.. - Confirm: (a) sample unmodified, (b) every remaining hunk is real config, (c) no reference URL hunk left, (d) no config line changed.
- Show the user the final diff and the survival list of their config before finishing.
Example: hyprland.lua (2026-09)
Concrete classification from a real run:
- Reference (applied verbatim from sample): wiki URLs restructured from
Configuring/Basics/...andConfiguring/Advanced-and-Cool/...to lowercaseconfiguring/core/...etc. — 15 comment hunks (monitors, autostart, env-vars, permissions, variables, tearing, animations, workspace-rules, dwindle/master/scrolling layouts, devices, binds, window/workspace-rules). - Config (kept, remain in diff):
fileManager = "thunar",menu = "wofi --show drun", the activehl.on("hyprland.start")block (fcitx5, waybar, hypridle, hyprsunset),hl.env("XCURSOR_THEME", "Adwaita"), hyprsunset+500/-500binds, the SCREENSHOTS block. - Ambiguous → ground truth:
dampeningvsdampingon a springhl.curve. Found upstream PRhyprwm/Hyprland#15993— a deliberate rename ("It IS a breaking change") merged Aug 29 2026 with a compat shim accepting the olddampening. Installed version here washyprland 0.56.2(Aug 5 2026), old enough thatdampingwas not yet recognized. Decision: keepdampening— works on the installed release and remains accepted after upgrade; stays as a legitimate config diff.
Related skills
Stores durable facts in a categorized, plain-markdown vault on disk, alongside your agent's built-in memory.
Find why your productivity system keeps failing, then apply the smallest fix — capacity math, bottleneck routing, durable local notes.
Write, debug, and tune Playwright specs with locator strategy, trace diagnosis, and CI-aware timeouts.
Join a video meeting as an AI bot with voice, avatar, and screenshare across four operating modes.
Run Git operations — commits, branches, merges, rebases, conflict resolution, and recovery — with safety rules enforced.
More from j3ffyang
Browse all skillsPolish a technical blog draft into an 800–1000 word, 3–4 section zh-CN article, preserve technical terms/code, and generate consistent hero + per-section image prompts when the user asks to polish and translate a blog with images.
Generate video from supplied image using Veo-3.0 async API. Gemini vision analyzes image, Veo creates video via predictLongRunning. All outputs in ~/.openclaw/workspace/tibetanProc/ with yymmddHHMM prefix.
Read a plain Markdown file (e.g. 260321_openclawConfig.md), generate header and section images using the default OpenClaw image model and API key from ~/openclaw/.env, create a new *_img.md copy with embedded PNGs, and move all generated images into an img/ folder next to the original file.
Polish a technical blog draft into a 1000–1200 word, 4–5 section en-US article, preserve technical terms/code, and generate one consistent hero image prompt.
Help the user draft, review, and polish blog posts in Markdown, keeping the original languages and informal tone.
Generate a daily AI news newsletter from fresh web sources. Use when the user asks for a current AI digest, AI news roundup, curated newsletter, or daily AI...