docType: guide scope: repo status: current authoritative: true owner: skills language: en whenToUse: "When installing, updating, or using the Tiangong AI reusable skills repository." whenToUpdate: "When install commands, target agents, scope behavior, environment variables, or available skill guidance changes." checkPaths:

• AGENTS.md
• .docpact/config.yaml
• README.zh-CN.md
• .claude-plugin/**
• "*/SKILL.md" lastReviewedAt: 2026-08-17 lastReviewedCommit: c26f4b17d8e50cd04267a1d86ff9d3ad9a07039a

Tiangong AI Skills

Repository: https://github.com/tiangong-ai/skills

Use the skills CLI from https://github.com/vercel-labs/skills to install, update, and manage these skills.

Install the CLI

npm i skills -g

Install

• List available skills (no install):

  npx skills add https://github.com/tiangong-ai/skills --list
  

• Install all skills (project scope by default):

  npx skills add https://github.com/tiangong-ai/skills
  

• Install specific skills:

  npx skills add https://github.com/tiangong-ai/skills --skill tiangong-auto-research --skill tiangong-kb-sci-search
  

Target agents and scope

• Target specific agents:

  npx skills add https://github.com/tiangong-ai/skills -a codex -a claude-code
  

• Install globally (user scope):

  npx skills add https://github.com/tiangong-ai/skills -g
  

• Scope notes:
  • Codex is a universal agent: project scope uses ./.agents/skills, and global scope uses $HOME/.agents/skills. CODEX_HOME does not change the skills@1.5.22 global destination.
  • Claude Code project scope uses ./.claude/skills; global scope uses $CLAUDE_CONFIG_DIR/skills when set, otherwise $HOME/.claude/skills.
  • Other agents have their own directories; inspect the exact path reported by the pinned skills CLI rather than deriving ~/<agent>/skills by analogy.

Install method

• Interactive installs let you choose:
  • Symlink (recommended)
  • Copy

Update and verify

• List installed skills:

  npx skills list
  

• Check for updates:

  npx skills check
  

• Update all skills:

  npx skills update
  

Environment Variables

Environment requirements live with each skill. Before using a skill that calls an external service, read that skill's references/env.md when present. npx skills add installs or links skill files; it does not provision language runtimes or execute post-install hooks. When a skill provides a locked runtime bootstrap and smoke command, run those explicit steps from its own instructions.

Auto Research External Skill Setup

Use npx skills directly for ordinary Skill management. For an Auto Research workspace, prefer the CLI's guarded setup layer: it still uses the exact pinned skills CLI underneath, while also binding source commits, tree hashes, destinations, license choices, safe credential bindings, and audit state. The Wizard lets ordinary users enter each selected key with hidden TTY input; named environment variables and bounded stdin/password-manager input remain explicit alternatives. Start with the read-only catalog or the guided Wizard:

REVIEWED_BOOTSTRAP_CLI_VERSION=X.Y.Z # replace with one reviewed exact stable release
npx --yes --package "@tiangong-ai/cli@$REVIEWED_BOOTSTRAP_CLI_VERSION" -- \
  tiangong-ai research setup catalog \
  --workspace /absolute/path/to/workspace --json
npx --yes --package "@tiangong-ai/cli@$REVIEWED_BOOTSTRAP_CLI_VERSION" -- \
  tiangong-ai research setup \
  --workspace /absolute/path/to/workspace

For repeatable non-interactive provisioning, run research setup init first. It creates a no-overwrite schema-v2 .tiangong-research/setup.yaml template plus setup.env.example. The YAML explicitly lists every current catalog Skill, credential, and setting, including disabled optional entries; the env example lists every credential variable with an empty placeholder. After the user reviews the current catalog, enabled states, licenses, models, pricing, checks, and confirmations, bare research setup detects only that workspace-local YAML and bypasses the Wizard. The optional real setup.env must be owner-only; disabled credentials stay empty and a non-empty disabled value is rejected. The removed v1 declaration and all invalid or incomplete declarations fail closed; they never trigger an interactive fallback or parent-directory search. Setup returns success only after every selected dependency, provider live check, Policy compatibility check, and independent reviewer smoke reaches complete readiness. Use explicit research setup wizard to choose the interactive path. See tiangong-auto-research/references/setup.md and tiangong-auto-research/references/env.md.

The bootstrap version is an explicit new-workspace choice, never latest, a tag, or a range. After apply creates runtime-lock.json, the installed orchestrator's bundled resolver runs exactly that locked version for all workspace operations.

The catalog also offers the tiangong-auto-research workflow orchestrator, default-baseline Brave internet evidence, optional Tiangong SCI/document/paper companions, and optional Anthropic or PPT Master post-closure authoring Skills. The workspace can be any user-selected directory. Every entry is external, separately licensed, pinned, and explicitly confirmed/selected; nothing is bundled or installed by a research package. See tiangong-auto-research/references/setup.md and external-skills.md. For PPT creation, prefer PPT Master; Anthropic PPTX remains compatible and may be selected alongside it when its workflow fits the task.

For a top-journal goal, the orchestrator includes a conservative Research Policy template pack for article type, field, journal class, project brief, and four independent final-review roles. The CLI Policy Wizard copies the selected Markdown into the research workspace for human review; it reports when generic defaults remain, requires explicit approval of the exact content hash, and invalidates approval after any edit or expiry. Exact-journal readiness requires current official guidance and substantive human customization. These gates can produce a reviewable submission candidate, never a promise of editorial acceptance.

Before discovery, the current native Codex or Claude host must also provide a closed, target-specific scientific design. The CLI validates and freezes the design, then enforces independent research-design, real-record evidence-construct, and pilot-methods reviews before discovery, acquisition, and analysis respectively. It reserves the complete early/final-review and revision lifecycle, requires reapproval for every authoritative recovery generation, and exports a portable audit directory containing exact formal evidence rather than host-local pointers. The CLI remains the deterministic control plane; it never launches a nested producer to invent the science. See tiangong-auto-research/references/publication-policy.md and tiangong-auto-research/references/scientific-design.md.