Skill detail

tiangong-auto-research

Orchestrates open-ended, multi-source, evidence-backed research in Codex, Claude Code, WorkBuddy, or CodeBuddy hosts, with setup, review, and closure workflows.

MatchPossibleRecommended for WorkBuddy
Sourcetiangong-ai/​agent-skillsExternal source
Reported installs4Popularity signal only

Inspect before use

Read the source instructions and check the requirements before you install this skill.

Saved source preview

SKILL.md

The saved excerpt is a snapshot from when this skill was listed. The external source remains the complete and most current version.

---
name: tiangong-auto-research
description: Orchestrate open-ended, multi-source, evidence-backed research in the current native Codex, Claude Code, WorkBuddy, or CodeBuddy host, especially when an ancestor directory contains `.tiangong-research` or the user asks to research, investigate, build on prior outputs, compare evidence, form a conclusion, or produce a reviewed research artifact. Also use for setup, preflight, native producer execution, recovery, independent review, closure, and requests to report Auto Research bugs or suggest capabilities (反馈问题、提交 Issue). In an Auto Research workspace this Skill takes precedence over individual web, news, SCI, report, patent, download, or document Skills unless the user explicitly requests one isolated standalone operation outside the research workflow. Covers Chinese requests such as “研究一下”, “朝这个方向做一做”, “结合已有成果继续研究”, “查资料并形成结论”, and “系统梳理证据”.
---

# Tiangong Auto Research

## Route feedback before research

For requests to report a problem, suggest a capability, or prepare an Issue,
read [references/issue-reporting.md](references/issue-reporting.md) and follow
that reporting workflow. Reporting does not require a valid research workspace,
setup, or scientific-question approval. Return the report without entering the
research workflow unless the user also requests research execution.

## Own the accepted research task

The user supplies goals, material scope preferences and necessary permissions;
the native assistant owns operationalization, planning, evidence work, technical
choices, checks, integration and an honest completion judgment. Preserve an
accepted goal through the existing task intake; do not ask the user to author
its JSON or repeat the same contract. No goal yet means no invented task facts.

On “continue”, a new session or context recovery, first follow
[task recovery](references/execution-assurance.md#recover-the-task-before-choosing-local-work)
before choosing another local experiment. An unchanged valid accepted question does
not need repeated approval; apply the question gate to a new substantive request
or new biased constraint. Project instructions installed on disk are not proof
that an already-running host loaded them; honor reported restart/compatibility
requirements and user overrides.

## Gate the research question before acting

Before any CLI, browser, search, database, or file operation, inspect the
user's research request. Continue when it leaves the result open to evidence.
Pause and require a rewrite when it assumes the conclusion, asks to prove a
predetermined position, requests only supporting evidence, excludes contrary
evidence, or treats an unsupported causal or normative judgment as fact.

When paused:

1. Do not call tools or begin setup.
2. Identify the problematic assumption in one concise sentence.
3. Offer one testable rewrite that preserves the intended topic, scope,
   population, geography, and time period.
4. Wait for the user's explicit confirmation or edited question.

Do not reject a controversial topic or directional hypothesis merely because
it has a position. It may proceed when null results, alternative explanations,
and counterevidence remain testable. Refuse requests to fabricate, conceal, or
misrepresent evidence.

For an existing managed directory, use the bundled resolver for ordinary CLI operations.
An explicitly reviewed CLI upgrade uses the candidate entrypoint in
[references/setup.md](references/setup.md), then returns to the activated lock.
Resolve `AUTO_RESEARCH_CLI` from this loaded Skill's absolute directory; do not
guess a global Skill path. The resolver accepts only the CLI package and exact
stable version from the regular non-symlink runtime lock, or from the immutable
setup plan while installation is still pending or blocked:

```bash
AUTO_RESEARCH_CLI=/absolute/path/to/installed/tiangong-auto-research/scripts/research_cli.mjs
node "$AUTO_RESEARCH_CLI" --workspace /absolute/path/to/workspace -- research --help
`
Read the full source on GitHub (opens external page)
Context

Related work