Serenity.skill: A Supply-Chain Research Workflow for AI Agents Focused on Chinese Equities

The Problem: Moving from Hot Topic to Research Priority

When a market theme like AI semiconductors, robotics or power equipment dominates financial news, the difficulty is not finding coverage but filtering it. Most retail investors can sense the heat; few have a repeatable process for deciding which tier of the supply chain is actually constrained, which companies sit closest to that constraint, and which are primarily benefiting from label association.

Serenity.skill packages a documented research path, attributed in the README to the publicly observable methodology of the analyst known as Serenity (@aleabitoreddit on X), into an installable Agent Skill. When you invoke it, the agent is instructed to start from a hot topic, decompose it into supply-chain layers from downstream demand through chip/device, equipment, materials, packaging and infrastructure, identify the layers with the fewest suppliers, the longest qualification cycles and the hardest capacity expansion, and then rank candidate stocks and fund directions by how close they sit to those bottlenecks.

The workflow ends with a prioritised research list plus explicit notes on missing evidence and risks, not a trading recommendation.

How the Skill Structures an Agent Research Session

The core document is SKILL.md, which the agent client loads when you invoke serenity-skill. The skill instructs the agent to work in five stages: read the hot topic to find the real demand driver; decompose the theme into downstream demand, system integration, chip/device, equipment, materials, packaging and infrastructure; identify bottleneck layers using criteria such as low supplier count, long validation periods, capacity expansion difficulty, strict customer qualification and high material purity requirements; map candidate stocks and fund directions to those layers; and verify each claim against public disclosures, exchange filings, earnings reports, earnings calls, regulatory/project documents, patents, industry standards, credible media and professional analysis.

The skill references several companion documents in the references/ directory, including deep-research-workflow.md, evidence-ladder.md, market-source-playbook.md and risk-and-compliance.md. Two worked examples ship in the examples/ directory: a real-research run on five A-share AI semiconductor companies covering public filings through 2026-09-14, and a challenge analysis of a specific CPO company's customer and capacity claims. A research template in assets/thesis-template.md organises findings into supply-chain position, confirmed facts, missing evidence, profit and valuation, alternative routes and invalidating conditions.

The README states that strong conclusions should be grounded in filings, exchange documents, earnings calls and credible analysis. Social media content including posts from Serenity is described as suitable for generating leads, not for final judgements.

Installing Serenity.skill in Codex and Claude Code

The repository requires an agent client that can load Skill files and that provides its own live search, browser or filing-retrieval tools. Research does not depend on Python; the Python 3 dependency is only for the structure-validation script.

Clone or download the repository first:

git clone https://github.com/muxuuu/serenity-skill.git
cd serenity-skill

For a user-wide Codex installation:

SERENITY_DIR="$HOME/.agents/skills/serenity-skill"
mkdir -p "$SERENITY_DIR"
cp -R SKILL.md LICENSE references assets examples agents "$SERENITY_DIR"/

For a user-wide Claude Code installation:

SERENITY_DIR="$HOME/.claude/skills/serenity-skill"
mkdir -p "$SERENITY_DIR"
cp -R SKILL.md LICENSE references assets examples agents "$SERENITY_DIR"/

To restrict the skill to a single project, replace the target path with the project's absolute path under .agents/skills/ or .claude/skills/. After copying files, open a new agent session and verify the skill is found before running a real research query. The README notes that Codex CLI 0.147.0 has been confirmed to load the skill and complete company-claim analysis on offline materials, while Claude Code directory and package structure have been checked but model invocation has not yet been end-to-end tested.

Maintainers can run the structure-check script from the repository root:

python3 scripts/validate_skill.py .

This command checks the skill name, description and directory layout. It does not evaluate the quality of any investment conclusion.

A sample prompt from the README:

用 serenity-skill 深度调研现在 A 股 AI 半导体产业链。
请联网查公告、财报、问询函、互动易、招投标、环评/能评、专利、客户认证和财务质量，
先排产业链层级，再给出通常 3–5 个值得优先研究的标的；证据不足时可以少给，
并说明卡住的环节、产业链位置、证据、排序理由和主要风险。

What the Agent Produces and What It Does Not

The README provides a template output structure that illustrates the intended result. The agent presents two to three research directions ordered by evidence strength, lists companies with their supply-chain position and the specific bottleneck rationale for each, identifies one or more directions it is deprioritising and explains why (typically: narrative is strong but order evidence, profit realisation or customer qualification is not yet confirmed), and closes with three next verification steps specifying a company, a document type (earnings report, announcement, filing) and the exact metric to look for.

The scope is deliberately bounded. The skill handles research prioritisation, evidence assessment, supply-chain positioning and risk flagging. It does not execute trades, access brokerage accounts, produce earnings-per-share estimates, confirm current stock prices or promise that any identified company will appreciate. The README repeats this boundary several times and frames it as a design choice: the agent handles the research, the human makes the decision.

The examples directory includes a CPO company challenge that illustrates the evidence-ladder approach, checking mass production status, customer identity, revenue quality and profit booking separately and marking each as confirmed, partially confirmed or unconfirmed. This level of granularity is available when the agent's connected data tools can retrieve the underlying filings.

Limitations and Cases Where It Is the Wrong Tool

Serenity.skill has no data layer of its own. It provides research instructions; the quality of the output depends entirely on the agent client's ability to retrieve live filings, exchange announcements and earnings materials. An agent running without internet access or filing-retrieval tools will produce structurally correct but factually unsupported output.

The methodology is tuned for Chinese A-share themes with public-disclosure ecosystems around exchange filings, regulatory documents and analyst calls. The README examples focus on AI semiconductors, CPO (co-packaged optics) and robotics supply chains within that market. Applying the skill to markets where the public disclosure conventions differ, for instance US small-cap technology or emerging-market industrials, will require the user to judge whether the evidence-ladder criteria translate.

End-to-end verification is incomplete. The README explicitly states that Codex CLI 0.147.0 has been tested with real offline materials, but Claude Code model invocation had not been fully tested as of the last documentation update. Users of other compatible agent clients must rely on the client's own documentation for skill-loading behaviour.

The skill does not produce numeric scores or synthetic ratings. Research prioritisation is explained in prose and evidence terms. Users who want a quantified output or a screener with numeric filters will need to build that layer themselves.

Maintenance and Licence

The last push to the repository was on 2026-09-16. The project is not archived. There are no GitHub releases; the main branch is the delivery vehicle. The skill ships under the MIT licence, which permits unrestricted use, modification and redistribution provided the copyright notice is retained. The research content in the references and examples directories and the methodology itself are not separately licensed: users should consult the repository's top-level LICENSE file for the full terms.

The CHANGELOG.md and CONTRIBUTING.md files are present in the repository, as is a SKILL.md that defines the skill's interface for agent clients. The README instructs users who are upgrading from an older version to move the old installation directory outside the skill search path before copying the new version, because a direct overwrite does not remove retired files and multiple copies in the same search directory may cause duplicate loading.

Editorial conclusion

Serenity.skill suits investors who already follow AI, semiconductors, robotics or cleantech themes and want an agent to do the first layer of structured research before they read any company filings. It is not appropriate for anyone expecting buy/sell signals or real-time data: the README is explicit that the skill provides research support only and that final trade decisions remain with the user. Before deploying it, verify that your agent client (Codex CLI 0.147.0 or later is documented as tested; Claude Code directory structure has been confirmed but end-to-end model invocation has not) can load and invoke external skills, and confirm that the agent has live search or filing-retrieval tools attached, because the skill itself ships no data feeds.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. muxuuu/serenity-skill on GitHub
4. README