Human Edge Agent Skills

Human judgment where generic AI is weakest: messy voice, social tact, ambiguous decisions, real-world experiments, and safe shipping.

English · 简体中文

AI is increasingly good at generating possibilities. The scarce part is still human: understanding a relationship, repairing noisy speech from context, deciding what matters this week, translating across fields, preserving a real voice, testing ideas against reality, and knowing when a result has enough proof to show the world.

This repository packages those recurring workflows as portable SKILL.md folders.

Start with four flagship Skills

Each Skill has one job, a trigger-rich description, and an independent install path. The rest of the repository is the supporting catalog—not eighteen equally loud claims.

The 18 Skills

Two voice-native Skills that work together

voice-dump-to-todo and vibe-to-spec deliberately share a context-aware ASR repair method, but have different destinations:

• voice-dump-to-todo → “What should I do?” → actions, decisions, dependencies, due dates, interactive HTML.
• vibe-to-spec → “What should the software do?” → scope, user flow, data/state, constraints, acceptance criteria, implementation order.

Both can use recurring named entities from the active conversation/project context to repair obvious ASR errors. They auto-normalize only high-confidence corrections; ambiguous names remain visible instead of being guessed.

Plugins

Five of the 18 canonical Skills are also packaged into four focused, self-contained plugins. The skills/ directories remain the source of truth; packaged copies are generated and hash-checked.

Each directory has a root Agent Plugins 1.0 manifest for portable clients such as Cursor and a Claude Code manifest under .claude-plugin/. No plugin adds hooks, agents, commands, MCP servers, or platform-specific behavior.

Current Codex can discover all four from the checked-in repository marketplace:

codex plugin marketplace add FAIRY123456789/human-edge-agent-skills --ref main
codex plugin list --marketplace human-edge-skills --available --json
codex plugin add voice-to-work@human-edge-skills

For a local clone, replace the GitHub source in the first command with .. The 2026-09-09 Codex runtime record covers discovery, installation, all five packaged Skills, uninstall, and clean reinstall. A marketplace listing is not implied.

Claude Code can add this repository as a marketplace and install one plugin:

claude plugin marketplace add FAIRY123456789/human-edge-agent-skills
claude plugin install voice-to-work@human-edge-skills

For a local clone, run claude plugin validate ., add the clone with claude plugin marketplace add . --scope local, and install with --scope local. Cursor's official local-development route is to copy one complete plugin directory into ~/.cursor/plugins/local/<plugin-name>, reload the window, and confirm its Skills under Customize. These files are packaging and submission preparation; they do not claim a marketplace listing or acceptance.

Install one standalone Skill from the monorepo

GitHub CLI 2.90+ supports browsing and installing a specific Agent Skill from a repository:

GitHub CLI:

gh skill preview FAIRY123456789/human-edge-agent-skills vibe-to-spec
gh skill install FAIRY123456789/human-edge-agent-skills vibe-to-spec --agent claude-code --scope user

For another compatible host, use the host supported by your current CLI or copy the individual skill directory into that agent's skills location.

The deployment arc: fit → ship → prove

The original lightweight safe-shared-vps-deploy concept has been deduplicated into the stronger aliyun-fullstack-deploy workflow. It retains the shared-server protection rules and adds release packaging, canary testing, atomic promotion, deeper verification, and rollback tooling.

• small-server-fit asks whether the app and runtime fit a low-cost server before purchase or deployment.
• aliyun-fullstack-deploy ships to an authorized Alibaba Cloud ECS or comparable Linux VPS through evidence gates.
• deployment-proof turns the redacted result into an honest public case study and career evidence.

Why one repository instead of 18 repositories?

The public entry point stays simple: one GitHub link on a resume or personal website, one star/watch/fork surface, one release history, and one place for issues. Individual Skills still have their own folder, SKILL.md, human-facing README, keywords, deep link, and can be installed individually from the same repository.

The repository includes a release strategy in docs/MONOREPO_AND_DISCOVERY.md. The recommended rule is: start as a monorepo; spin out only a Skill that independently earns enough users or contributors to justify its own lifecycle.

Current AI-podcast route

deep-talk-to-podcast includes a dated platform selector for source-grounded Audio Overviews, editable AI podcast production, and high-quality generated voices. Because product names and features change quickly, the Skill tells an online agent to verify current official pages before recommending a service.

Validation

pip install pyyaml
python scripts/validate_skills.py
python scripts/sync_plugin_packages.py --check
python scripts/validate_plugin_packages.py

Seed behavior evals live in evals/cases.json.

The NVIDIA SkillEvaluator Tier 1 audit is recorded in docs/NVIDIA_TIER1_REPORT.md, with the candidate classification and fix log. Raw JSON is retained under eval-results/.

Codex packaging and runtime validation are recorded in the Codex runtime report. Cross-platform state is summarized in Cross-Runtime Marketplace Status. The Plugin Submission README provides one copy-ready source for Claude and Cursor form fields; platform-specific records remain available for OpenAI, Cursor, [Cla