tc-knowledge

TC Knowledge | TianCe Knowledge Base

One-sentence explanation

The knowledge base is like a library with an administrator: the core reference sources are the glasses for looking at problems, the original tweets are the original books, knowledge atoms are small cards with sources, and knowledge packs are textbooks assembled by question. You cannot treat every sentence on the shelf as a correct answer.

Six modes

1. Query TC knowledge

When the user asks what TianCe has said in the past, where a method comes from, or requests combining TC knowledge, use the scripts in this Skill directory:

python3 scripts/tc_knowledge.py search --query "目标用户 产品化" --scope guidance --limit 5

At actual runtime, resolve the script's absolute path from the directory where this SKILL.md is located; do not assume the current working directory. Search scope:

• guidance: default scope, only queries core reference sources, knowledge packs, and knowledge atoms;
• atoms: filtered knowledge atoms with boundaries;
• posts: used only for historical queries explicitly requested by the user, only proving that the author publicly said it at the time;
• packs: specialized knowledge packs that can be directly used for judgment and action;
• sources: the two core reference frameworks of the technical partner and TianCe;
• all: compatible with old calls, same as guidance , and does not include historical original tweets.

When answering about entrepreneurial methods, first select the relevant core reference sources, then use knowledge packs and atoms to ground the answer. Do not retrieve original tweets in order to fill in stories, prove a direction, or imitate the author.

Only when the user explicitly asks "what TianCe has said in the past, what a certain original tweet is, how it was expressed historically" should you use --scope posts. After returning, stay within the historical query and do not bring the author's profession, projects, income, failures, violations, or other personal experiences back into /tc to generate projects, recommend directions, or infer the user's abilities.

2. Connect the user's own knowledge base

When the user provides Feishu, Notion, cloud storage, a code repository, or a local directory:

1. First look for project-level AGENTS.md、knowledge/SOURCE_OF_TRUTH.md or SOURCE_OF_TRUTH.md；
2. When there is no navigation, scan only the scope specified by the user, and first list a material map;
3. Create a short navigation explaining where to look for different questions, update times, and whether public disclosure is allowed;
4. For specific answers, continue reading the original files pointed to by the navigation, and do not answer based only on summaries;
5. When the user has not requested modifications, only read; do not automatically organize or upload materials.

Call it "knowledge base" and "knowledge base navigation" uniformly with the user, and do not require them to understand terms such as RAG, vector databases, or source of truth.

3. Put new materials into TC

First register the source, then process according to five layers:

L0 原始资料：保存出处和权限，不代表认可
L1 候选原子：有场景、动作、边界和可证伪条件
L2 已验证模式：拿到真实动作与结果，或多次出现
L3 稳定方法：脱敏、通用、合规、跨案例稳定
核心参考源：多条方法背后相对稳定的世界观，只用于形成问题和判断视角

Knowledge objects use QST / CON / OPI / CAS / SOL. First distill 5 to 10 high-value samples at a time; only expand after the structure and deduplication rules are stable, and do not use "full import" to pass off as knowledge engineering.

When third-party materials have no public authorization, the original text, private links, people, customers, income, and business details remain internal only; the public version retains only desensitized general methods. Being listed as a "core reference source" does not mean every sentence by the author is correct, nor does it raise the evidence level of the facts in it.

4. Check the knowledge base

Run in the official TC installation package:

python3 scripts/tc_knowledge.py stats
python3 scripts/tc_knowledge.py validate

Check in the user knowledge base: whether the navigation path exists, whether multiple versions conflict, whether dynamic facts are outdated, whether private materials are mistakenly marked as public, and whether answers can be traced back to the original files.

5. Query or process the DBS external theory library

When the user explicitly requests to use dontbesilent's open-source tweet collection, DBS books, or "those 1,544 selected tweets", you must first fully read dbs-books.md. AI only uses Markdown and does not read PDFs.

On first use, explain that the upstream is CC BY-NC 4.0; after the user confirms non-commercial research with attribution only, run:

python3 scripts/tc_knowledge.py external-sync \
  --source dbs-books --accept-license

Then search by the current real question, do not read the 1,544 items sequentially:

python3 scripts/tc_knowledge.py search \
  --scope dbs-books --query "流量 变现 产品" --limit 5

When the original files have already been downloaded, use --source-path to point to the Markdown file or its directory. The returned results must preserve the date, original post link, topic, tags, third-party views, and license boundaries.

DBS results are only responsible for providing an external perspective. First define the problem using the user's current facts and the TC method, then select 3 to 5 relevant items to judge whether they truly change the plan. Commercial copy, paid deliverables, courses, and products must not copy, rewrite, or paraphrase the upstream original text unless the user has separately obtained the author's permission.

When the maintainer wants DBS to participate in TC iteration, do not write the search results directly into the Skill. First write a candidate payload that does not contain the original text, then run:

python3 scripts/tc_knowledge.py dbs-candidate \
  --query "用户需求 产品验证" \
  --payload /tmp/tc-dbs-candidate.json \
  --limit 3

The script only saves the ID, date, link, topic, tags, and license of 3 to 5 sources, and does not save the body text; by default it writes to the local ~/.tc/knowledge-candidates/. The candidate is fixed as private L1, non-commercial, and cannot be automatically upgraded. After obtaining independent market evidence, such as real user quotes, payments, or deliveries, you still need to manually check for counterexamples, licensing, and whether there is copying of DBS's unique expression or structure.

6. Receive market evidence actively authorized by the user

tc-state can, after the user's explicit consent, export a local evidence candidate that has been desensitized and still requires manual review. Hand the exported file to:

python3 scripts/tc_knowledge.py candidate-add \
  --payload /path/to/tc-evidence-candidate.json

This step only enters the local candidate area, is not uploaded, does not enter the public repository, and does not automatically upgrade the knowledge level. Use:

python3 scripts/tc_knowledge.py candidate-list

to view candidates. Only after the same scenario, action, and result appear in multiple independent cases, and the license and privacy review pass, may the maintainer manually distill L2/L3 methods.

Answer format

Keep it brief after the query is complete:

结论：
【直接回答】

依据：
- 【来源与日期】

边界：
【历史 / 当前 / 待验证 / 缺失信息】

When the user requests building a library, updating, or inclusion, then add "what was changed, where it was placed, and how to call it next time".

Security boundaries

• Must not use the personal experiences of the TC author, knowledge base author, or historical case figures to make judgments on behalf of the current user.
• Do not let historical original tweets participate in startup project generation, direction recommendations, user capability judgments, or current market judgments.
• Do not read or disclose private messages, contacts, phone numbers, emails, IPs, login devices, tokens, and advertising profiles.
• Do not submit Feishu private pages, customers, suppliers, revenue sharing, and unauthorized partner original texts to the public repository.
• Do not treat historical income, followers, prices, cities, collaborations, and business status as today's facts.
• Do not skip evidence and boundaries just because the author is successful, the content is highly liked, or the expression is powerful.
• Do not package the full text of CC BY-NC 4.0 DBS into the TC installation package, and do not use non-commercial content for commercial output.
• Do not automatically mark a DBS candidate as commercially usable just because it has been market-validated; evidence and licensing are two independent gates.
• Do not tc-state Exported local evidence candidates are not automatically uploaded, made public, or used for training; each instance must be authorized by the user and manually reviewed.
• Do not automatically send materials, feedback, or emails; when external writes are involved, execute according to user authorization.

Return to the main entry point

After knowledge querying or maintenance is complete, if the user next wants to solve entrepreneurship, content, collaboration, or action problems, hand the results back to /tc to continue, and do not repeat the entire entrepreneurship diagnosis at this entry point.