127 lines
5.7 KiB
Markdown
127 lines
5.7 KiB
Markdown
---
|
|
name: auto-research
|
|
description: Research uncertain questions with an explicit, user-approved web search or ChatGPT consultation, then present options and wait for implementation approval.
|
|
category: automation
|
|
risk: critical
|
|
source: self
|
|
source_type: self
|
|
date_added: "2026-07-09"
|
|
author: zyu51
|
|
tags: [research, chatgpt, playwright, browser-automation, decision-support, chinese]
|
|
tools: [claude, playwright]
|
|
license: MIT
|
|
---
|
|
|
|
# Auto-Research Skill
|
|
|
|
## Overview
|
|
|
|
When implementing tasks, Claude Code can encounter uncertainties — design choices, algorithm details, API usage, or best practices. This skill provides an explicit-consent research path, presents findings, and waits for user approval before writing code.
|
|
|
|
The skill supports web research and an optional ChatGPT consultation. It never sends
|
|
conversation context, files, browser state, or credentials to a third party without the
|
|
user's explicit approval of the exact, redacted text.
|
|
|
|
## When to Use This Skill
|
|
|
|
- User asks a question where multiple valid approaches exist
|
|
- Claude is uncertain about algorithm details or API usage
|
|
- Design/architecture choices need comparison
|
|
- The user explicitly asks to search the web or consult ChatGPT and approves the proposed query
|
|
|
|
## How It Works
|
|
|
|
**Step 1: Propose the research boundary** — State the source to use, the exact query or
|
|
redacted prompt, whether any local/workspace text would leave the machine, and the likely
|
|
cost. Wait for the user to approve that exact boundary.
|
|
|
|
**Step 2: Research** — After approval, use web search or a browser session the user has
|
|
explicitly selected and authorized. Use a pinned, user-configured browser automation
|
|
connector; do not install packages automatically, use `@latest`, or access browser cookies,
|
|
other tabs, saved passwords, or sessions.
|
|
|
|
**Step 3: Present** — Distill findings into concise options with sources, presented to the user.
|
|
|
|
**Step 4: Await Approval** — Do NOT write code until the user says "go ahead" or picks an option.
|
|
|
|
**Step 5: Implement** — Once approved, execute with confidence.
|
|
|
|
### Explicit ChatGPT Consultation
|
|
|
|
Do not treat `?`, `??`, or another shorthand as consent. First propose a minimal prompt,
|
|
for example: `请评估这个已脱敏的方案的正确性、完整性和可改进之处：<text>`.
|
|
Explicitly identify every piece of text that would be sent. Only after the user confirms
|
|
the exact prompt may you open the selected ChatGPT session, submit that prompt, and present
|
|
the response. Do not include conversation history by default.
|
|
|
|
Redact secrets, personal data, proprietary code, customer data, and internal URLs before
|
|
proposing the prompt. If safe redaction is not possible, do not submit it.
|
|
|
|
### Browser Automation Boundary
|
|
|
|
If browser automation is necessary, the user must separately authorize the selected browser
|
|
profile and connector version. Restrict the session to the consultation tab. Do not inspect,
|
|
reuse, export, or rely on cookies from other tabs or profiles.
|
|
|
|
## Examples
|
|
|
|
### Example 1: Design Question with GPT
|
|
```
|
|
User: PyTorch 中自定义 ADMM 优化器怎么设计？
|
|
Claude: 我可以搜索公开资料，或将以下已脱敏问题发给 ChatGPT：
|
|
“如何设计 PyTorch 自定义 ADMM 优化器？请比较可行模式。”
|
|
不会发送工作区文件或对话历史。是否允许？
|
|
User: 允许发送这段文字
|
|
Claude: [Opens only the authorized consultation tab, submits the approved prompt]
|
|
Claude: GPT suggests approach A with these pros/cons. Proceed?
|
|
User: 行
|
|
Claude: [Implements code]
|
|
```
|
|
|
|
### Example 2: Web Search
|
|
```
|
|
User: ?? ADMM convergence criteria best practices
|
|
Claude: I can search public sources for the exact redacted query
|
|
“ADMM convergence criteria best practices”. No workspace files or conversation
|
|
history will be sent. May I send that text to WebSearch and fetch the results?
|
|
User: Yes, send that query
|
|
Claude: [WebSearch + WebFetch → finds Boyd et al. paper, extracts criteria]
|
|
Claude: Boyd recommends ||r|| < ε·max(||Ax||, ||Bz||, ||c||). Use this?
|
|
User: Yes
|
|
Claude: [Implements]
|
|
```
|
|
|
|
## Best Practices
|
|
- ✅ Always present findings to user before writing code
|
|
- ✅ Use `page.fill()` for instant text injection instead of `keyboard.type()`
|
|
- ✅ Ask for fresh approval before every external consultation
|
|
- ✅ Include sources in findings
|
|
- ❌ Don't skip research and write code speculatively
|
|
- ❌ Don't send context, files, or browser data because of a shorthand trigger
|
|
- ❌ Don't alter the user's browser profile or session state
|
|
|
|
## Limitations
|
|
- Requires a user-configured, pinned browser automation connector if browser consultation is used
|
|
- ChatGPT consultation is optional; use ordinary web search when it meets the need
|
|
- GPT response time varies (10-30s typically)
|
|
- Web search quality depends on available sources
|
|
- Does not replace expert domain knowledge — always let user make the final call
|
|
|
|
## Security & Safety Notes
|
|
- Obtain explicit consent for each third-party submission, including the exact redacted text
|
|
- Never access, export, or depend on cookies, saved passwords, or unrelated browser tabs
|
|
- Never submit sensitive credentials, tokens, proprietary code, personal data, or internal URLs
|
|
- Do not install or execute browser tooling from an unpinned package version
|
|
|
|
## Common Pitfalls
|
|
|
|
| Problem | Solution |
|
|
|---------|----------|
|
|
| ChatGPT shows login page | Let the user log in themselves; do not handle cookies or credentials |
|
|
| The prompt contains sensitive context | Redact it or use local reasoning instead |
|
|
| Browser automation is unavailable | Use web search or stop and ask the user for a different approved method |
|
|
|
|
## Related Skills
|
|
- @systematic-debugging — use when debugging Playwright interactions with ChatGPT
|
|
- @condition-based-waiting — use when waiting for GPT responses in the browser
|