Promptbase

是什么

Promptbase 是把提示词（Prompt）当代码管理的工程实践：版本化、可测试、可灰度。 用它的效果是：提示词的每次改动都有 diff、有评测、有回滚，不再靠口口相传。

怎么用

1. 把每个提示词独立成文件，让 Git 历史记录每次表达调整的来龙去脉。
2. 为每个提示词配套 golden 测试集，让效果变化在合并前就被自动发现。
3. 用模板变量隔离稳定文案与动态参数，让提示词可以被多场景复用。
4. 通过灰度发布在小流量上验证新版本，让大盘风险被控制在可承受的范围内。
5. 把每次 A/B 实验的结论写回提示词的 changelog，让团队沉淀经验而非重复造轮。

架构图

flowchart LR
  提示词草稿 --> 版本仓库
  版本仓库 --> 评测套件
  评测套件 --> 灰度发布
  灰度发布 --> 生产流量
  生产流量 --> 经验回写

PromptBase Skill Publishing

Patterns for writing Claude Code skills that pass PromptBase review and sell well.

When to Activate

• Writing a new skill intended for sale on PromptBase
• Reviewing an existing skill before submission
• Adapting a rejected skill after a "too specific" or "too simple" ruling
• Writing listing copy — title, description, examples, setup instructions
• Deciding whether a skill idea is worth submitting at all

The Core Test: Would a Stranger Pay For This?

Before writing a single line, answer three questions:

If any answer is red, fix the concept before writing the skill.

Rejection Reasons and Fixes

Too Specific (most common rejection)

What it means: The skill only works for users of a niche platform, private SDK, or internal framework. The audience is too small to justify listing it.

Signs your skill is too specific:

• It imports from a private package (from mycompany.lib import ...)
• It references internal tool names, ports, config files, or env vars no outsider would know
• It combines more than 3 niche ideas (e.g. "LangGraph + Agentex + custom state machine")
• The target framework has fewer than ~5k GitHub stars

Fix: generalize the layer

REJECTED:  Temporal + Agentex ADK + adk.state + adk.messages + FastACP
APPROVED:  Temporal + Python temporalio SDK — standard patterns any developer can use

Strip the platform-specific layer. Keep the hard, transferable patterns underneath.

Too Simple / Guessable

What it means: A buyer could recreate the skill by reading the title and thinking for 30 seconds. The value isn't there.

Signs your skill is too simple:

• The entire skill is one pattern or one code snippet
• It documents official defaults (e.g. "how to create a FastAPI route")
• It has no decision tables, no red flags, no non-obvious gotchas
• Word count is low AND there are no patterns that require explanation

Fix: add depth

• Add a decision table (when to use X vs Y)
• Add a Red Flags section (7–10 anti-patterns with explanations)
• Add the non-obvious gotchas — things that only experience teaches
• Add a pre-ship checklist

No Use Case

What it means: The skill has no practical application for real work — it's a demo, experiment, or overly academic topic.

Fix: Anchor every section to a real scenario a working developer would face. If you can't name a concrete task the skill helps complete, the concept needs rethinking.

Scope Rules

The 3-niche rule: PromptBase rejects skills that combine more than 3 niche ideas. A "niche idea" is any tool, pattern, or concept that a developer might not already know. Standard Python, SQL, and REST are not niche. LangGraph, Temporal, and a custom state machine pattern together are three niche ideas — borderline.

Writing the Listing

Title

# BAD — platform-specific, tiny audience
Agentex Temporal Agent Skill

# BAD — too vague, no value signal
Python Async Skill

# GOOD — specific problem, broad audience
Crash-Proof AI Agents: Temporal Workflow Skill for Claude Code

Formula: [Outcome the buyer wants]: [Technology] Skill for Claude Code

Description (keep it under 100 words)

Structure:

1. One sentence on the problem it solves
2. 4–6 bullet points on what Claude knows after activation
3. One line on format and installation

Your Temporal workflows break in ways that are invisible until production.
This skill teaches Claude the patterns that prevent it.

- Workflow vs activity split and determinism rules
- Retry policies, timeouts, and heartbeats
- State management across replays
- Signal handling and `wait_condition`
- 7 red flags that silently break Temporal agents

One `.md` file. Drop in `~/.claude/skills/` — activates automatically.

When to Use (activation description)

Write this as a single sentence starting with "Use when" that describes triggering conditions — not a summary of topics:

# BAD — topic summary, not a trigger
Use when working with Temporal, workflows, activities, signals, and state.

# GOOD — triggering condition
Use when building or debugging Temporal workflows in Python — structuring
activities, enforcing determinism, handling retries, or diagnosing replay failures.

Examples (2 required)

Each example is a realistic user message + a sample Claude response showing the skill in action.

Example 1 (shown as preview on store page): Pick the sharpest, most impressive demo — the problem that would make a buyer think "I need this right now."

Example 2: Show a different use case from Example 1. Don't demonstrate the same pattern twice.

# BAD example pair — same pattern twice
Example 1: "How do I add a retry to my activity?"
Example 2: "How do I set a timeout on my activity?"

# GOOD example pair — different concerns
Example 1: "My workflow stops receiving signals after the first message."
Example 2: "Can I call httpx directly inside my workflow?"

Setup Instructions

Keep it to 3 steps maximum:

1. Download `skill.md`
2. Place it at `~/.claude/skills/<skill-name>/skill.md`
3. Restart Claude Code

No MCP servers, env vars, or dependencies required.

If the skill requires external tools (MCP server, API key), list them explicitly. Buyers who can't set up the skill will leave bad reviews.

Skill Body (the SKILL.md content)

PromptBase asks for the body only — everything after the YAML frontmatter. The frontmatter is generated on their end. Paste starting from the first # heading.

Pre-Submission Checklist

• Skill uses only publicly available, open-source tools — no private SDKs
• Topic is not reproducible from the title alone — real value is in the patterns
• Combines at most 3 niche ideas
• Has at least one decision/comparison table
• Has a ## Red Flags section with 6–10 domain-specific anti-patterns
• Has a ## Checklist section with 8–12 pre-ship items
• Title follows [Outcome]: [Tech] Skill for Claude Code formula
• Description is under 100 words and leads with the problem
• description frontmatter starts with "Use when" and describes trigger conditions only
• Two examples show different use cases, not the same pattern twice
• Setup instructions have 3 steps or fewer
• Skill body pasted without frontmatter (PromptBase generates it)
• No references to internal systems, private repos, or proprietary env vars

Red Flags

• Private SDK imports in examples — any from yourcompany.internal import ... immediately signals the skill is platform-locked; reviewers reject these without reading further; replace with the standard open-source equivalent
• Title that gives away the entire skill — if the title is "Use retry_policy=RetryPolicy(maximum_attempts=3) in Temporal", a buyer can reproduce it without purchasing; titles should name the outcome, not the solution
• Description longer than 150 words — PromptBase listing descriptions overflow on mobile and store pages; buyers skim; if you need 200 words the concept isn't focused enough
• Examples that demonstrate the same pattern — two examples showing retry configuration are one example, not two; reviewers flag this as low variety; each example must address a distinct concern
• Setup instructions that require a paid API key or private access — buyers who hit a wall during setup leave negative reviews; if the skill requires external services, state it prominently in the listing before purchase
• "Use when" description that summarizes topics instead of triggers — "Use when working with Redis, caching, pub/sub, and streams" is a topic list; "Use when choosing a Redis data structure, implementing rate limiting, or debugging slow commands" is a trigger list; only triggers activate the skill correctly
• Submitting a skill that is a thin wrapper around official docs — if every pattern in the skill appears verbatim in the framework's own documentation, it adds no value; the skill must contain non-obvious patterns, gotchas, and decisions that experience teaches