Copy Brain

copy-brain

Copy a public figure specified by the user into a callable thinking skill.

The core is not imitating tone, but replicating the "brain". Tone and wording are only surface-level; what we truly want to capture is how this person thinks, how they weigh trade-offs, and how they make decisions under uncertainty. The final skill should enable the agent to reason out the answer this person would likely give on a new problem using their thinking style, rather than reusing things they have already said. Path convention: The directory containing this SKILL.md is the skill root directory (hereafter ). All paths below are **relative to **. Before running any script, first cd into `` (the directory containing this SKILL.md); output/template paths are resolved relatively in the same way.

Workflow

Step 0 / 1: Check the output directory and determine the persona

1. List the .md files under /output/ (ignore .gitkeep; ignore stray intermediate files like .json, see Step 4.1).
2. If skill files already exist: list them for the user and ask "Do you want to reuse one of the existing skills, or create a new persona skill?".
   • If they choose to reuse → read that file and complete the user's subsequent request based on it, then end this workflow.
3. If the directory is empty: directly ask "Whose thinking would you like to copy?".

Once you have the persona's name, do a lightweight confirmation: is the name unique (if there are namesakes, ask the user to add field/nationality), and determine a slug for the filename (pinyin or English, lowercase, hyphenated).

---

Step 2: Check and explain the API Keys

This skill can use three data sources (recommended but not required; skip whichever is missing). Ask the user to configure the corresponding API Keys in their computer's system environment variables using the variable names in the table below:

Explain the table above to the user (purpose + recommended setup). How to configure (after setting, you must reopen the terminal for it to take effect):

# Windows PowerShell (persist to user environment variables)
setx TAVILY_API_KEY "your_key"
setx SCRAPEBADGER_API_KEY "your_key"
setx REDFOX_API_KEY "your_key"

# macOS / Linux (add to ~/.zshrc or ~/.bashrc)
export TAVILY_API_KEY="your_key"
export SCRAPEBADGER_API_KEY="your_key"
export REDFOX_API_KEY="your_key"

Check the current configuration:

python scripts/check_keys.py

If dependencies are missing, install first: pip install -r requirements.txt (run in ``)

• The script signals that a service is not configured with EXIT_NO_KEY (exit code 2) + a stderr message.
• Even if none of the three are configured, you can continue using just the agent's built-in web search + web extraction tools, only with narrower coverage. Do not force the user to configure them.
• The Tavily script only exposes 3 search parameters: query (required), --search-depth (optional, basic/advanced/fast/ultra-fast, default basic), --time-range (optional, day/week/month/year). For full text, use the built-in web extraction tool on the result URLs.

---

Step 3: Search background → determine scenario

1. First use the agent's built-in web search to do a quick background check on the persona (identity, field, notable achievements, active platforms). If the built-in results are poor (stale / multilingual / poor matching), use Tavily as a fallback:

python scripts/tavily_search.py "{persona} bio background"
# When you need higher relevance or a time window:
python scripts/tavily_search.py "{persona} bio background" --search-depth advanced --time-range year

1. Summarize the quick-check highlights to the user in 2-4 sentences, then ask about the scenario: "In which scenario would you like to think/decide using this persona's thinking style?" and offer a few candidate examples (tailored to the persona's traits), emphasizing thinking and judgment rather than tone.

After the user confirms the scenario, record scenario and its slug.

---

Step 4: Deep search → draft → confirm → save

4.1 Collect deeply and broadly (multiple angles, multiple keywords). Prioritize first-hand original platform sources—as long as a Key is configured, use the scripts below as much as possible, treating the persona's own posts/notes/articles as the primary corpus; supplement general background/commentary with the built-in search (Tavily fallback when poor): Every script supports -h to view its parameters.

# ① First-hand platform sources (primary)—X platform (persona is active on X and ScrapeBadger is configured)
python scripts/scrapebadger.py profile {x_username}
python scripts/scrapebadger.py tweets {x_username} --pages 3
# You can also precisely search their tweets by topic (advanced operators):
python scripts/scrapebadger.py search "from:{x_username} {scenario_keywords}" --type Latest --pages 2

# ① First-hand platform sources (primary)—Xiaohongshu / WeChat Official Accounts (Chinese personas, RedFox configured)
python scripts/redfox_xhs.py search "{persona}" --pages 2
python scripts/redfox_gzh.py search "{persona} {scenario_keywords}" --sort _4 --pages 2
# After a hit, use detail / work to pull the full text of a single piece to get the complete reasoning:
python scripts/redfox_xhs.py detail --id {workId}
python scripts/redfox_gzh.py work {workUuid}

# ② General background, views, controversies, recent developments—built-in search first, Tavily fallback when poor
python scripts/tavily_search.py "{persona} {scenario_keywords}" --search-depth advanced
python scripts/tavily_search.py "{persona} views quotes interview" --time-range year

Every fact/view/piece of reasoning must keep a source link; especially preserve original excerpts that reveal "why he thinks this way".

• Read the full text: both the built-in search and Tavily return summaries/snippets; to read the full text, use the agent's built-in web extraction tool on the result URLs. If a platform post fetched by a script only has a summary, use redfox_*.py detail/work or extract the original link to get the full text.
• Do not save intermediate JSON to disk: the scripts print JSON to the terminal stdout by default, and the agent can just read the terminal output. Do not use --out, and do not save .json or other intermediate collection files under output/. output/ only holds the final persona skill .md (see 4.3). If you accidentally generate a .json under output/ or an output/output/ subdirectory, delete it after saving the .md.

4.2 Generate the draft: apply the structure of /templates/persona_skill_template.md, filling it in for the chosen scenario. Present the draft to the user for preview, explain the data sources and collection time, and proactively point out areas that are uncertain / lack evidence. 4.3 Save after confirmation: once the user agrees, save to:

/output/{people_name_slug}-{scenario_slug}.md

After saving, tell the user the file path and note: this file is itself a persona skill with frontmatter and can be reused.

Quality and Ethics Requirements

• Factual claims need sources; distinguish between "facts" and "stylistic simulation". Do not fabricate quotes or falsify data.
• Do not use this to impersonate the actual person for fraud, defamation, or misleading purposes; do not include private/non-public information.
• Search with the utmost rigor. If the information found is insufficient to copy the public figure's thinking, stop generating the draft and inform the user.