Start by using it: Zhang Zala follow-builders installation and first reading
follow-builders retrieves the public information sources maintained by the author, then hands them to your Agent to organize according to language and reading preferences. For first use, it is recommended to first see a result with original links in the current chat, and then configure automatic push separately; this makes it possible to distinguish whether the problem lies in content acquisition, model organization, or delivery.
Below is organized according to the upstream commit verified on 2026-10-25 313d901913d007ee6014cf2f622d00e6086e812f. The subsequent historical share and comments remain preserved; “install and use immediately” and “no key needed” do not mean there are no model fees, push credentials, or always-on scheduling conditions.
Manual installation entry point in Claude Code
You need an already usable Claude Code, Git, Node.js/npm, and network access to GitHub and the central feed. The following paths come from the upstream installation instructions. If the directory already exists, first check the current contents and version; do not clone again or overwrite. Preserve your own modifications before updating.
git clone https://github.com/zarazhangrui/follow-builders.git ~/.claude/skills/follow-builders
cd ~/.claude/skills/follow-builders/scripts
npm install
After giving the skill to your own Agent to read, you can say: “Set up follow-builders. First display this issue’s summary in Chinese in the current chat, retain source links, and do not enable automatic push for now.” The upstream README uses /follow-builders as the entry point,SKILL.md also describes /ai; the specific slash command depends on whether the host recognizes that skill. If it is not recognized, first check the installation location and the host’s loading result; do not treat the command name itself as successful installation.
OpenClaw users can install from the corresponding entry point in the upstream README and verify the channels supported by the version used; do not treat Claude Code’s directory directly as a universal path for all hosts.
What counts as a successful first result
First confirm that the output actually contains this issue’s content, source links, and feed time, rather than merely generating an “already set up” response. Then check the language and summary style. If a certain type of source is empty, inspect the corresponding feed content and error messages; do not equate an empty array with the entire system being broken, and do not fabricate entries.
In this actual read of the three feeds at the fixed commit, all returned HTTP 200: the X feed had 14 authors, the podcast feed had 1 item, and the blog feed had 0 items; no feed errors were reported. This is a content snapshot of that version, not a guarantee of daily quantities. A dependency installation check was also completed under Node.js 23.6.1/npm 10.9.2, with installation lifecycle scripts disabled; the Agent’s first-run guidance, scheduled tasks, and Telegram/email delivery were not actually tested, so these stages cannot be recorded as accepted.
What conditions are still missing for automatic push
Content reading uses the author’s central feed and does not require users to apply for their own X or subtitle service key. Model operation still depends on your Agent service; when choosing channels such as Telegram or email, corresponding channel configuration and credentials are required. Do not paste these credentials into public articles or commit them into repositories.
According to the current SKILL.md, OpenClaw can use its always-on scheduling and messaging channels; for other hosts to deliver on a schedule, they separately require continuously running scheduling and an available delivery method. Merely installing the skill in the terminal or manually reading one issue does not prove that it will send automatically after the terminal is closed. Before formally enabling it, at least check that: the task is actually created, the scheduled time and time zone are correct, execution records show success, and the target channel actually receives the same content.
If results are always empty, first look at the feed date and errors; if it can only run manually, first check scheduling; if the backend shows success but nothing is received, check the delivery channel. Changing the local information source list also does not necessarily change the crawling scope of the author’s central service; these two matters must be kept separate.
References:Fixed-version README,Fixed-version SKILL.md.
Historical share and comments: original content remains preserved
Author:zara Zhang Zala (GitHub: zarazhangrui) Current materials: 1 item (2026-05-18 video special-topic share, see 2026-05-18-follow-builders-skill-fenxiang) Core skill:follow-builders, open-sourced; on the user’s local machine
~/.claude/skills/follow-builders/Installed
When to open it
- When selecting AI information sources / wanting to build an AI briefing / wanting to turn “carefully selected sources + automatic push” into a distributable product
- Wanting to understand whether, when a Skill depends on external APIs, users should get the key or the author should centralize it this Skill engineering branch
- For a cross-author synthesis of Skill engineering topics, see Claude Skill Engineering (Cross-Author Topic Synthesis) §6.5
1. Underlying philosophy: Follow builders, not influencers
Information sources for learning AI should be first-hand:
- First-hand = AI industry practitioners (researchers, entrepreneurs, developers) ← original viewpoints
- Second-hand = KOL / self-media interpretation ← regurgitate
zara earlier built a website for this (a builder list + recommended podcasts) as a compilation; later users gave feedback that “following one by one takes time and I can’t understand English,” so the website was upgraded into a skill.
2. follow-builders skill product form
Every morning it automatically pushes an AI briefing to Feishu (via OpenClaw) or other channels. The content is divided into two blocks:
| Block | Content | Carefully selected scope |
|---|---|---|
| Builder tweet summaries | OpenAI / Cursor / Google Labs CEO + Box CEO + well-known AI investors + zara’s personal picks “my personal intelligence” | |
| Podcast | AI substantive podcast summaries | Shows that frequently interview frontline labs / outstanding entrepreneurs / researchers; the original videos can be reviewed |
Adjustable parameters: push time / frequency / language (Chinese / English / Chinese-English bilingual) / length (summary → in-depth → full transcript) / style.
3. Key iteration v1 → v2: Centrally fetch, locally Remix
v1 failure: users’ Agents crawl themselves
Each user’s Agent calls Supadata (YouTube transcripts) + the official X API itself → almost everyone gives up at the first step:
- X API paid + complex configuration
- Most users are unwilling/unable to obtain an API key
v2 success: Centrally fetch, locally Remix
Author side (zara centralization):
daily script → calls APIs (zara pays) → fetches latest builder content → writes into the JSON data source
User side (skill installed in OpenClaw / any Agent):
skill trigger → fetches zara’s central JSON → locally remixes according to user preferences (language/length/style) → pushes
| Dimension | v1 user self-crawling | v2 central + local remix |
|---|---|---|
| User barrier | High (self-configured API key, paid X) | Zero (install and use immediately) |
| API cost | Borne by users | Borne by the author (controllable scale) |
| Content quality | Determined by the prompt written by the user | Carefully selected by the author, quality assured |
| Personalization | Both content and form are personalized | Unified content, form locally remixed |
| Value proposition | Tool | Careful selection + tool |
Core quote:“The greatest value of this skill is some very high-quality information sources after my careful selection and filtering” — the value is not in the distribution mechanism, but in the careful selection itself.
4. Spillover of the methodology: scenarios for reuse
Any product form involving “externally maintained data by the author + user-side skill invocation” can apply this pattern:
- Industry briefings(this skill case): select N high-quality information sources, the author maintains crawlers + JSON, and the user’s skill provides personalized presentation
- Professional knowledge bases: the author maintains an authoritative knowledge-base JSON, and the user’s skill performs RAG according to the scenario
- Dataset distribution: data requiring continuous updates, such as stocks / housing prices / recruitment / paper metadata
- Paid API relay: the author pays for fetching + distributes free of charge (can be combined with a subscription model for monetization)
5. Installation and triggering
- Go to GitHub:
github.com/zarazhangrui/follow-builders - Send the skill to your OpenClaw or any Agent for installation
- onboarding: The Agent proactively asks about preferences such as “push time / frequency / language” and writes them into the user configuration
- Trigger: the Agent schedules automatic runs, or use
/ai(the skill name corresponding to Claude Code isfollow-builders)
Related
- Cross-author Skill engineering topic compilation:Claude Skill Engineering (Cross-Author Topic Synthesis) §6.5
- With dontbesilent AI self-media course / 13-lesson systematic methodology §L13 (Skill engineering vs. general methodology) complementary
- Same topic (AI information source management):newtype knowledge system and full evolution of Life OS / follow-builders local skill
History
- 2026-05-18: First release, based on the follow-builders skill topical sharing video (Zara Zhang Zala, publicly released on 2026-05-18)
Additional comments from March to June 2026: distinguish content updates, adaptations, and operating conditions
This section is based on the March 19 post and the crawled author replies and real-person comments; it does not treat the automatic summary from the comments section's “Ask a Question” as a verbatim transcript of the video, nor has it rechecked the repository's operating status.
On March 21, the author explained that the JSON is automatically updated daily; on April 22, she suggested that people who want to add new accounts to follow should fork it (copy the repository to create their own version) and then ask the agent to modify it.Here, “being able to change the list” does not mean that changing only the local list will alter the crawling scope on the author's server: Users in the same comment thread confirmed this question with positive feedback, but no closed-loop answer was seen. When additional information sources are needed, one should distinguish subscribing to an existing feed from taking responsibility for crawling, authorization, and scheduling oneself.
Historical comments recorded issues such as empty content, exhausted API quota, feeds stopping updates, and long posts being truncated; the author replied on May 7, 14, and 31 respectively that they had been updated or resolved. This shows that centralized crawling lowers the personal configuration threshold while also concentrating update dependency on the maintainer. Comments cannot prove that a fault still exists today, nor can they prove that nothing was ever missed historically. For knowledge-base purposes, checking date coverage and original IDs is more reliable than looking only at the latest briefing.
The author's “not needed” reply regarding an X/Supadata key pertains to using the central data source she provides; it does not mean that any self-built crawling is free or requires no credentials. Users still have conditions related to models, operating devices, and push channels. Some people reported that they could only run it manually, while others reported scheduling failures; the author's interaction suggestion to “directly state what time each day to send” cannot replace verification that a task has been created and successfully delivered.
When to use: when you want to retain the author's information filtering and then convert it into Chinese or an HTML reading format. Counterexample: if you need a complete historical archive for your own industry, you cannot rely solely on this latest feed; user claims in the comments such as “saving 60% of tokens” have not been empirically tested and are not listed as promised benefits.