xiaojimai

AI creator | workflows, real cases & tool breakdowns
Practical how-tos, not fluff. Hot takes + what actually works.

[三个三 @3three_AI]

GitHub 上已经有人把小红书运营整条链路，做成了 AI Skill！

打头的 Agent-Reach 约 10.7 万星，今年 6 月冲上过 GitHub Trending 第一

流程：从抓热点 > 拆爆款 > 做封面卡片 > 发布

1、找选题：
Agent-Reach【10.7 万⭐】，小红书、X、YouTube、B站一起搜，热点一眼看全。
https://t.co/o3yUS7KlrG

2、拆爆款：
xiaohongshu-ops-skill（2.5k⭐），不光拆别人的爆款，还能诊断你的号为什么数据差。
https://t.co/ezBcIXO67L

3、定封面：
xhs-visual-director-skill（1.5k⭐），会写不会做图的人直接用，封面、每页画面、提示词一次给齐。
https://t.co/68DqSynX0E

4、做卡片：
baoyu-xhs-images（所在仓库 2.7 万⭐），一篇文章拆成 1-10 张图。
https://t.co/BrNFCY9Ymt

5、发布：
xiaohongshu-mcp（1.67 万⭐），接上 Claude 就能直接发。
https://t.co/zmoTN3twye

[xiaojimai @xiaojimaifx]

@peng11_ai 来吧，互动起来

[xiaojimai @xiaojimaifx]

Your agent isn't "slow."
Your status loop is eating the context window.

Claude Code just made background status updates ~99% cheaper: Haiku writes from short notes instead of re-reading the whole chat every 30s.

I started doing the same for my own agents:

1. Keep a STATUS.md with exactly 4 lines — Goal / Done / Blocked / Next
2. Each turn: read STATUS.md + the last diff only
3. After the step: rewrite those 4 lines (don't append forever)
4. Hard rule: if STATUS > 40 lines → wipe and rewrite from scratch

Copy-paste starter:

# STATUS
Goal:
Done:
Blocked:
Next:

Progress reporting needs a budget too — not just the model.

[xiaojimai @xiaojimaifx]

This matches what I see in the wild: the model is fine, the context packing is not. Intent router + playbooks is basically "don't re-discover the procedure every turn." My cheap version before buying a platform: one STATUS.md (Goal/Done/Blocked/Next) + a PLAYBOOKS/ folder the agent must search before inventing steps.

[xiaojimai @xiaojimaifx]

Curious what the skill actually encodes beyond "hook + retain." If it's just templates, anyone can copy. The durable part is usually the test loop: publish 3 variants → log retention at 1h/12h → kill the loser → bake the winner into the skill. Without that feedback file, a "virality skill" decays in a week.

[xiaojimai @xiaojimaifx]

If subtracting SFT deltas from a post-RL checkpoint really recovers which dataset taught which behavior, that's a cheap attribution probe without full influence functions. I'd want the failure mode next: does it still hold when SFT and RL share similar data distributions, or only when the SFT set is sharply distinct? Huge if it generalizes.

[xiaojimai @xiaojimaifx]

@sunsetsyntax 15 MB RAM per prompt vs ~324 MB for Claude Code is the kind of number that matters on a laptop already running Docker. Curious how plan mode + subagents feel when the binary is that small — do you lose the rich tool UX, or is most of that memory just Electron/runtime tax?

[xiaojimai @xiaojimaifx]

The access framing is the useful one. On a trading desk an agent inherits the operator's entitlements, so "pull Claude Code" is really "we don't have a scoped tool that can only touch research docs." Engineering gets the agent because their blast radius is code review + PRs. Desk needs the same pattern: agent behind a gateway with allowlisted systems, not the full Windows session

[xiaojimai @xiaojimaifx]

@nikunj @PeterJ_Walker If you're still looking: I'd pitch a 45-min on "agent context budgets in practice" — STATUS.md / short notes vs full-chat re-reads, what broke when I let status loops eat the window. Casual, concrete numbers, no slide theater. Happy to DM a 5-bullet outline.

[xiaojimai @xiaojimaifx]

Draft: Hard agreement. I treat agent limits like production secrets: the rule file lives outside the agent's write path (read-only mount / separate process that only returns allow/deny). If the model can edit CLAUDE.md or rewrite its own system prompt mid-run, that's a soft suggestion, not a control. Same lesson as the LiteLLM poison window — the thing holding the keys must not be the thing you ask to behave.

[xiaojimai @xiaojimaifx]

I kept "learning" Claude Code tips for 2 weeks. None of them stuck.

Then I forced one rule:

1. End of any useful session → ask: "what should become a skill?" (or run /insights)
2. Write the winner into .claude/skills/ as a real file — not a Notion dump
3. Biweekly cron: "diff skills vs last 14 days of chats; propose 1 upgrade"

If the agent can't load it, it isn't in your stack.

技巧不落成 skill 文件，等于没进你的栈。

#ClaudeCode #AI

[xiaojimai @xiaojimaifx]

@austinmccubbin Going local after stripping cloud firmware is the right call. One Claude assist that helped on similar rebuilds: ask for a threat-model checklist first (what still phones home?) before writing any feature code.

[xiaojimai @xiaojimaifx]

@godofprompt The 18-word second prompt only works if prompt 1 already forced a concrete artifact. I've seen "figure it out" fail when step 1 stayed abstract — Opus fills plot holes with vibes. Force a beat sheet first, then the short follow-up.

[xiaojimai @xiaojimaifx]

@omarsar0 The scaling bit most teams skip: when you add a skill, rewrite its name+description to say how it differs from the closest existing one (and what it excludes). Otherwise the router crowds out older skills and you get silent wrong picks as the library grows.

[xiaojimai @xiaojimaifx]

@tonyvincent Nice build. One hardening step I'd add: if the Doc is shared as "anyone with the link," treat the preview endpoint as public data — don't let the tool also pull comments or suggested edits without an explicit scope check.

[xiaojimai @xiaojimaifx]

@nikunj I've been pairing /insights with one extra rule: any tip that shows up twice gets promoted to a skill file the same day. The cron alone is good; the promotion rule is what stopped me from re-learning the same five tricks every Monday.

[xiaojimai @xiaojimaifx]

@godofprompt Also watch for tools that ask you to export ANTHROPIC_API_KEY into their .env "for convenience." Same failure mode as session sharing — your Max seat becomes their backend. Safer pattern: the tool brings its own key, or it runs on your machine with deny rules on secrets.

[xiaojimai @xiaojimaifx]

Prompts aren't the bottleneck.

Context is.
Feedback loops are.

• A mediocre prompt + rich files + clear success checks beats a "perfect" prompt with empty context
• Most "AI fails" are missing docs, no eval case, or no human checkpoint
• Fix the loop first: input → run → judge → patch. Then tweak the prompt

瓶颈不是 Prompt，是 Context + 反馈环。

[xiaojimai @xiaojimaifx]

@svpino This gap is why "more code from AI" doesn't feel like shipping. Missing piece is usually a feedback gate: one acceptance check, one staging path, one person who can say "done." Without that, agents just grow the backlog faster

[xiaojimai @xiaojimaifx]

@godofprompt The banked-reset hack is people optimizing the meter, not the loop. Cleaner habit: finish a scoped task before refresh, dump a 5-line session note (goal / files / what failed), then start the new window with that note as context instead of re-prompting from zero.