Building AI-powered automation skills, productivity tools, and knowledge systems for individual developers and AI-native workflows.

I create production-ready automation skills across multiple AI platforms (OpenCode, Hermes, OpenClaw, Claude), focusing on:

• Skill-based AI orchestration — turning workflows into reusable, composable skills
• Multilingual content automation — English ↔ Chinese/Traditional Chinese
• Developer productivity tools — Chrome extensions, CLI utilities, API integrations
• AI platform experimentation — deeply hands-on with OpenCode, Hermes, OpenClaw, Claude, and OpenRouter

All code prioritizes clarity over cleverness: simple, readable, portable, and easy to fork.

⚡ OpenCode — the growing collection, auto-published to ClawHub

My OpenCode skills live in ai-thoughts/.opencode/skills/ and publish to ClawHub automatically on every push to j3ffyang — no local installs, one owner account. The one to open first:

• asd-ste100 — write and audit agent instructions in ASD-STE100's rules: one action per sentence, condition first, one name per item. Ships a deterministic linter for sentence length, semicolons, banned words, and passive voice.

The rest of the collection — article translation, Traditional-Chinese twin sync, Astro sync, and resize-for-banner for Twitter/X headers only (LinkedIn and blog headers take the original image) — is in the repo linked below.

Timeline is editorial — edit it directly in scripts/portfolio_template.md, then regenerate.

• Oct 26: Published "Replace Cursor with a Minimal VSCodium AppImage" — English plus Simplified and Traditional Chinese twins, a 16:9 infographic, and a 5:2 banner, and synced to the Astro blog; a 3-function editor (git-tagged tree, Vim keys, Mermaid + local PlantUML) in one AppImage, with the honest 795 MiB Electron footprint and a one-step uninstall
• Oct 23: Published "Where Does Brave's Leo AI Keep Your Data?" — English plus Simplified and Traditional Chinese twins, a 16:9 infographic, and a 5:2 banner; the local store turns out to be decryptable with Chromium's public peanuts key on a machine with no keyring
• Oct 23: Restricted resize-for-banner to Twitter/X only — LinkedIn and blog headers accept the original image — and codified "register an article only after it is committed and pushed" in AGENTS.md, which removed three uncommitted rows from the manifest
• Oct 20: Published "The Rules, Not the Words: ASD-STE100 for Agent Instructions" — English plus Simplified and Traditional Chinese twins, a 16:9 infographic, and a 5:2 banner
• Oct 20: Added the asd-ste100 skill and a --stats summary mode to its linter; rewrote both working-rule AGENTS.md files to the same standard (18 and 15 linter errors to zero) and logged the diffs
• Oct 20: Added a portfolio_latest: false opt-out, so an article whose doc is not committed yet stays counted but off the front-page list; recorded a per-genre precision rule (approximate figures in essays, exactness in code and measurements)
• Oct 7: Refreshed the portfolio timeline and skills list, and dropped the WIP encrypt-linux row from the article index until the articles are ready
• Oct 7: Published "What Should You Know to Work Well with an AI Agent?" in three versions — English, Simplified Chinese, and its first Traditional-Chinese -zh-hant.md twin
• Oct 7: Added the simplified-to-traditional skill — -zh-hans → -zh-hant twin sync with OpenCC (s2twp), permanent cross-links, and --check; published to ClawHub 1.0.0
• Oct 7: Renamed the translation skill translate-to-chn → translate-to-zh-hans and published it to ClawHub 1.0.0
• Oct 7: Standardized the Simplified-Chinese suffix as -zh-hans.md across AGENTS.md, the skills, and the README generator's desc_zh check — legacy -chn.md files kept as-is until explicitly migrated
• Oct 4: Finished and published "Agent session persistence in OpenCode" (file-dated Sep 20, completed today) — bilingual (EN + ZH) with its own infographic — and refreshed the already-published "What Should You Know to Work Well with an AI Agent?"; both registered in the article index and portfolio

• Sep 30: Added the security-writeup-bad-effects skill, a verify_md_boxes.py box-diagram verifier, and a global AGENTS.md "Host safety" rule; cross-posted the Brisova malware analysis to Medium and X (5:2 banner, plus a platform-filter-safe rewrite of one exploit-describing line)
• Sep 29: Published "Anatomy of a Malicious GitHub Invite — The Brisova Build-Time Payload" (EN + ZH) with infographic; added the untrusted-code-safety skill and its global AGENTS.md "Untrusted code safety" contract
• Sep 27: Published "Choosing a Model for OpenCode via OpenRouter" with infographic and Twitter banner
• Sep 21: Published the rdr2-playthrough skill to ClawHub; reworked the portfolio pipeline so each account self-publishes its own README (no cross-account PAT or mirror)
• Sep 21: Expanded the portfolio Code Highlights with automation tooling (clawhub_publish.py, sync_profile.py, gen_readmes.py)
• Sep 20: Published "Extracting PowerPoint to Markdown Without Losing a Byte" (EN + ZH) with infographic and the pptx-extract skill
• Sep 17: Added the normalize-whitespace and sync-config-with-sample skills
• Sep 15: Added an auto-generated ClawHub skills table and switched READMEs to recency ordering
• Sep 15: Published the Chinese translation of the "create SKILL.md, AGENTS.md, and PERSONA.md on the fly" essay
• Sep 14: Published "Create SKILL.md, AGENTS.md, and PERSONA.md on the fly" with hero infographic and Twitter banner; renamed the ClawHub publish skill to publish-skills
• Sep 14: Refined the bold-highlights skill — sparser bolding, no list bolds
• Sep 2: Reworked the GPD dual-AMD eGPU article with mermaid diagrams, new image assets, and an Arch sign-off
• Aug 8: Added scripts/unwrap_md.py — auto-wraps article and skill prose to one-paragraph-per-line, preserving code fences, tables, and list nesting
• Aug 8: Published the AI agent collaboration playbook — what a real multi-repo ClawHub auto-publish project taught about AGENTS.md, SKILL.md, and project boundaries
• Aug 7: Published AMD iGPU + eGPU setup guide for GPD Win4 (bilingual EN/ZH) with verification commands and stable DRM symlinks
• Aug 4: Merged the three Chrome extensions (sum2chn, twitter2md, twitter-bookmark-summarizer) into a single chrome-extensions monorepo
• Early Aug: Published "Why OpenCode is the best AI agent setup for me" essay and the Ollama → llama.cpp local-LLM deep-dive (draft)
• Jul 31: Published the Hermes Agent skills/plugins cleanup guide (bilingual)
• Jul 21-22: Published Obsidian + Karpathy LLM Wiki local search guide and the "unknown unknowns" essay (bilingual)
• Jul 14: Added Brave browser privacy analysis (bilingual) to ai-thoughts
• Jul 14: Polished and committed two blog posts with bilingual README updates
• Early Jul: Releasing v1 of sum2chn Chrome extension (translation + summarization)

• Jun: Completed twitter2md Chrome extension and Node.js CLI tool
• Jun: Refactored twitterBookmarkSum to use popup-triggered summarization
• May: Released twitter2md for X post extraction as Markdown

• Apr: Published first Hermes Agent articles and tutorials in ai-thoughts
• Mar: Established openclaw-custom-skills repo on ClawHub
• Feb: Began deep-dive experimentation with OpenClaw and Hermes platforms

Automation tooling that keeps this portfolio, its READMEs, and its ClawHub skills in sync.

Repo: ai-thoughts | Language: Python | Purpose: Render PORTFOLIO.md from a template + articles.yaml + git log

def render(data: dict) -> str:
    out = TEMPLATE.read_text(encoding="utf-8")
    values = {
        "ARTICLE_COUNT": str(article_count()),
        "LATEST_ARTICLES": latest_articles(data),
        "LAST_UPDATED": last_updated(),
    }
    for key, value in values.items():
        out = out.replace("{{" + key + "}}", value)
    leftover = sorted(set(re.findall(r"\{\{[A-Z_]+\}\}", out)))
    if leftover:
        print(f"error: unfilled placeholders: {leftover}", file=sys.stderr)
        sys.exit(1)
    return out

Why it matters: Turns articles.yaml + git log into a fresh PORTFOLIO.md with zero hand-editing, and --check mode keeps the committed file honest in CI.

Repo: ai-thoughts | Language: Python | Purpose: Join hard-wrapped paragraphs into single-line paragraphs while preserving code fences, tables, and lists

def glue(a: str, b: str) -> str:
    a_end, b_start = a[-1:], b[:1]
    if CJK.match(a_end) and CJK.match(b_start):
        return ""          # 中文之间不加空格
    return " "

Why it matters: Article and skill prose is stored one-paragraph-per-line (auto-wrapped); this tool restores that convention across docs/ and .opencode/skills/ without touching structure — and --check flags any file that drifts.

Repo: ai-thoughts | Language: YAML + Python | Purpose: Publish .opencode/skills/ to ClawHub on every push — no local CLI

name: ClawHub Skill Sync
on:
  push:
    branches: [main]
    paths:
      - ".opencode/skills/**"
      - ".github/workflows/clawhub-skill-sync.yml"
  workflow_dispatch:

jobs:
  publish:
    if: github.event_name != 'pull_request' && github.repository_owner == 'j3ffyang'
    runs-on: ubuntu-latest
    permissions:
      contents: read
    steps:
      - uses: actions/checkout@v7.0.1
      - uses: oven-sh/setup-bun@<sha>
        with:
          bun-version: 1.3.10
      - name: Checkout ClawHub CLI
        uses: actions/checkout@v7.0.1
        with:
          repository: openclaw/clawhub
          ref: v0.23.3
          path: clawhub-source
      - name: Install ClawHub CLI dependencies
        working-directory: clawhub-source
        run: bun install --frozen-lockfile
      - name: Publish skills to ClawHub
        env:
          CLAWHUB_TOKEN: ${{ secrets.clawhub_token }}
        run: python3 scripts/clawhub_publish.py

Why it matters: Publishing runs the ClawHub CLI directly instead of the reusable workflow, so the five real statuses the CLI can return — unchanged, would-publish, submitted, published, pending-publication — are all treated as success. (The upstream reusable workflow only maps three of them, so a successful async publish showed up as a red ✗.) The repository_owner guard still stops the two remotes from double-publishing to one ClawHub account.

Repo: publish-skills skill | Language: Shell | Purpose: Confirm what's live on ClawHub via the read-only skills API

# What's actually live? Check latestVersion + owner.handle
curl -s -H "Authorization: Bearer $CLH_TOKEN" \
  https://clawhub.ai/api/v1/skills/<slug>

Why it matters: Turns "the workflow failed" into a verified fact — the API call proves a skill published even when the upstream workflow's status mapping reports a false failure.

Repo: ai-thoughts | Language: Python | Purpose: Drive the pinned ClawHub CLI to publish every .opencode/skills/* folder, treating all five real statuses as success

OK_STATUSES = {"unchanged", "would-publish", "submitted", "published", "pending-publication"}
...
status = json.loads(completed.stdout).get("status")
if status in OK_STATUSES:
    ok.append({"slug": target.name, "status": status})
else:
    failed.append({"slug": target.name, "status": status, "message": f"Unknown publish status: {status}"})

Why it matters: The upstream reusable workflow only maps three of the five statuses, so a successful async publish (pending-publication) surfaced as a red ✗. This treats all five as success and prints an ok/failed JSON summary the workflow can trust — while the repository_owner guard still stops the two remotes from double-publishing to one ClawHub account.

Status: legacy. The PORTFOLIO.md pipeline superseded this script; kept as a record. Repo: negtivSpace | Language: Python | Purpose: Recompute every derived number in the profile README (skills, repos, articles, Quick Links) from the nested repos

def apply(pattern: str, repl: str, desc: str, flags: int = 0) -> None:
    nonlocal text
    new, n = re.subn(pattern, repl, text, flags=flags)
    if n:
        text = new
        edits.append(desc)
    else:
        failures.append(desc)

for pattern, repl, desc in subs:
    apply(pattern, repl, desc)

Why it matters: Counts drift the moment a skill or article is added; one pass regenerates them, and the script exits non-zero if any expected section fails to match, so drift surfaces instead of shipping silently. Counting only git-tracked skill dirs keeps local leftovers from skewing the result versus CI.

Repo: ai-thoughts | Language: Python | Purpose: Rebuild README.md and README_zh.md from articles.yaml, with a --check mode for CI

if args.check:
    ok = True
    for target, generated in ((README_EN, en), (README_ZH, zh)):
        current = target.read_text(encoding="utf-8") if target.exists() else None
        if current != generated:
            ok = False
            print(f"stale: {target.name} is out of date")
    return 0 if ok else 1

Why it matters: Both language READMEs come from one manifest, so the English and Chinese indexes can't diverge, and --check makes CI fail on a stale README rather than letting it drift.

Languages: Python, JavaScript, TypeScript, Bash, YAML

AI Platforms:

• OpenCode — terminal-native AI coding agent; my daily driver
• OpenClaw — local agent orchestration
• Hermes Agent — autonomous workflow automation
• Claude Code — AI-native coding
• OpenRouter — unified LLM gateway

APIs & Services:

• Anthropic Claude (Sonnet, Opus)
• OpenRouter (100+ model access)
• Chrome Extension APIs
• Twitter/X API v2
• Web scraping & DOM manipulation

Tools:

• Git + GitHub
• GitHub Actions — CI/CD automation (YAML workflows)
• Node.js + npm
• Python 3.10+
• Markdown-first documentation

1. Skills > Projects — I think in workflows. Every tool becomes a reusable skill.
2. Multilingual by default — English + Chinese (Simplified + Traditional)
3. Code clarity wins — Readable Python/JS beats clever one-liners. Always.
4. Composition over bloat — Small, focused repos that work well together.
5. Learn in public — Detailed articles about what works (and what doesn't).

• GitHub: @negtivspace
• Blog: ai-thoughts — 51 articles on AI, privacy, and entrepreneurship
• Gists & Experiments: Personal gists

From ai-thoughts (most recent published articles):

1. "Health Records an Agent Can Audit" (Oct 2026) — A two-layer skill for AI-assisted review of personal and family medical records: append-only data, conflict protocol, change gate, evidence grading — and the honest list of what it still cannot do
2. "The Rules, Not the Words: ASD-STE100 for Agent Instructions" (Oct 2026) — ASD-STE100 Simplified Technical English for agent instructions — the standard's history and rules, three worked examples, a linter-backed before-and-after measurement of two rule files, and a skill that applies the rules without flattening essay voice; with a visual summary infographic
3. "Replace Cursor with a Minimal VSCodium AppImage" (Oct 2026) — Replacing Cursor with a VSCodium AppImage and three Open VSX extensions — git status tags in the file tree, Vim keys, Mermaid in the built-in preview and local PlantUML; why Cursor's forced agent and backend routing are the real problem, what telemetry surface remains, the settings that make the editor inert, and an honest 795 MiB Electron footprint with a one-step uninstall; with a visual summary infographic
4. "Where Does Brave's Leo AI Keep Your Data?" (Oct 2026) — Inside Brave Leo's local store — the single AIChat SQLite database, its schema for conversations/pages/search sources, why its at-rest OSCrypt encryption falls back to a hardcoded key without a keyring, what that structured local data could be used for, and why harvesting it team-wide is not worth it
5. "Unix Is Art" (Oct 2026) — A personal essay on the Unix philosophy as an aesthetic — its history from 1974 to 2004 (Thompson and Ritchie, McIlroy, Kernighan and Pike, Raymond), the ideas behind it and the critics of it, a remembrance of Dennis Ritchie, and what Unix made me
6. "Keeping secrets out of Git" (Oct 2026) — A short checklist for keeping secrets out of git — why a committed secret is already compromised, the .gitignore patterns to add, where to state the policy (README, AGENTS.md), and three commands to verify nothing is hiding; with a visual summary infographic

👉 See all 51 articles at ai-thoughts/docs

git clone https://github.com/negtivspace/ai-custom-skills
cd ai-custom-skills
# Follow README for your specific platform (Claude Code / Hermes / OpenClaw)

# Chrome extensions monorepo (twitter2md, sum2chn, twitter-bookmark-summarizer)
git clone https://github.com/negtivspace/chrome-extensions
cd chrome-extensions/twitter2md
npm install
npm run build  # or load the extension manually in Chrome

git clone https://github.com/negtivspace/ai-thoughts
cd docs
# 51 articles: AI platforms, privacy, solo entrepreneurship, technical deep-dives

Most repos are MIT License — see individual repos for details.

Last Updated: October 26, 2026 | Tracking: 6 active repos, 51 published articles, 20+ published skills