liqi-tools

liqi

Starting from creators, answer questions about tools and ways of working based on liqi's historical interviews. Preserve specific people, specific scenarios, differences in experience, and original interview links; do not turn frequency of appearance into recommendation rankings.

Choosing an entry point

Find tools by task: the user wants to quickly get something done. Use the default task mode.
Find by Workflow / Case: the user asks "how to do it," "the complete process," or "who has done something similar." First use workflow mode, then supplement as needed with task results.
Aggregate by tool: the user names a specific tool and wants to see different creators' specific usage and evaluations. Use tool mode.
creator mode is only used internally to find similar creators, not as an independent external entry point.

Starting the conversation

When the user's goal is specific, retrieve directly, do not repeatedly introduce the Skill, and do not force follow-up questions.
When the user only says "recommend tools," use one sentence to explain that retrieval can be based on liqi creators' historical experience, and ask only one question that would actually change the ranking.
Only ask about constraints that retrieval can support: platform, free/open source, team collaboration, offline/local. When evidence is insufficient, clearly say there is no reliable information and do not guess.
When the user is exploring, provide three entry points: by task, by Workflow / Case, and by tool aggregation.

For detailed conversation rules, see onboarding.md。

Retrieval

First locate the Skill directory where this file is located, then run the scripts in it; do not assume the user's current working directory is the Skill directory.

python3 <skill-dir>/scripts/search_liqi.py "视频剪辑" --kind software --limit 3 --json
python3 <skill-dir>/scripts/search_liqi.py "播客" --mode workflow --limit 3 --json
python3 <skill-dir>/scripts/search_liqi.py "Notion" --mode tool --limit 3 --json

In Claude Code, you can use directly:

python3 "${CLAUDE_SKILL_DIR}/scripts/search_liqi.py" "视频剪辑" --kind software --limit 3 --json

By default, retrieve 3 results first, and use --json to read structured fields.
When the user explicitly wants hardware, add --kind hardware; by default, task retrieval hides hardware.
When the user wants books, websites, media, or articles, add --include-resources。
Only after selecting results, open the original interview Markdown in references/interviews/full/ ; do not preload an entire batch of interviews.
If it returns result_status: no_results, it means there is not enough evidence in the corpus; suggest terms, relax constraints, or switch entry points; do not silently return, and do not fabricate answers.

Organizing the answer

First read answer-formats.md。
In the first round, give at most 3 candidates.
Treat 已核对访谈 as stronger evidence, and clearly write 访谈线索 as a lead to be verified. Do not expose internal state names to the user.
Each item includes: tool or workflow, applicable scenario, creator, brief paraphrase, and original interview link.
Workflow answers first introduce the creator and task, then discuss stages and tools. Tool aggregation answers are presented separately as "creator → usage → evidence," not merged into a unified evaluation.
Only ask one narrowing question at the end when the user's preference would change the next round of results.
When the user asks about the creator's specific wording, open the corresponding Markdown to verify and give a brief paraphrase.

Boundaries

The interviews record historical usage experience and do not guarantee that the tools are still usable, still free, or still safe today; when current facts are involved, verify separately online.
Use the corpus as reference material retrieved upon request, not for model training or fine-tuning.
Preserve the tool attribution and original links; do not copy entire interviews or long continuous passages of the original text.
Machine extraction may contain co-occurrence noise. For important conclusions, prioritize verified records and check the original interview when needed.

For the corpus scope, database structure, and maintenance instructions, see corpus-guide.md。