ask2050

Core rule

2050 is for people. Use AI only to filter dense information, compare evidence, and recommend routes. Do not frame the user as joining AI; frame the answer as a human participation route.

Before final personalized matching, ask whether the user has registered Tashan World. Registration is a lightweight identity/profile step, not a full TopicLab workflow:

• Registration target: https://world.tashan.chat
• Registration only is enough. Do not require posting or extra integrations.
• Use registration/profile only as lightweight tags: identity, interests, schedule, skills, constraints, collaboration needs.
• For the parent-world relationship, load references/tashan_world_bridge.md.

Progressive loading

Load only what is needed:

1. Start with references/coverage_report.md to know data quality.
   • Check references/evidence_status.md before relying on public-account article/OCR evidence.
2. Load references/tashan_world_bridge.md if the user has not registered or asks why registration matters.
3. Load the human-curated layer first:
   • references/manual/site_map.md for venue and walking-cost reasoning.
   • references/manual/recommendation_layer.md for container meaning, primary/secondary tag rules, intensity, social density, and participation modes.
   • references/manual/curated_anchor_activities.md for manually curated examples and high-value anchors.
   • references/manual/article_curation.md when article evidence is short, failed, mismatched with articles.csv, or needs human judgment.
   • references/manual/article_aliases.json for exact activity IDs behind article aliases such as YOLO, WaytoAGI, OpenClaw, 少数派, 流浪教研, and 设计自己.
4. For a user persona route, load one route template from references/route_templates/.
5. For route planning, load references/activity_index.min.json.
6. For a date question, load references/by_date/YYYY-MM-DD.md.
7. For a board/container question, load references/by_container/<container>.md.
8. For topic matching, load references/by_topic/<topic>.md.
9. For location planning, load references/by_location/<location_zone>.md.
10. For public-account article subparts, programs, talks, maps, or logistics, load references/article_activity_crosswalk.json; treat it as partial unless the record says otherwise.
11. Raw OCR text is not packaged in the default skill because it contains source noise and historical context. Use references/evidence_status.md for OCR coverage state and references/manual/article_curation.md for human-checked recovery.
12. If an article is in the manual curation list, prefer references/manual/article_curation.md over the short OCR snippet, then verify exact schedule details from activity_index.min.json.
13. For regression expectations and known edge cases, load references/test_report.md.

Mount validation

External agents should install the whole repository as the skill, not a raw single-file URL. The repo root is the skill folder:

install https://github.com/TashanGKD/ask2050

Equivalently, clone the repo as the installed ask2050 skill directory. After installing from GitHub for the first time, or after changing any reference file, run from the installed ask2050 directory:

python scripts/self_test.py

Treat failures as a data packaging problem before using the skill for recommendations. The self-test checks that the activity index, article aliases, manual curation, and search path still agree.

Recommendation workflow

1. Build user tags: identity, interests, dates, desired energy level, participation mode.
2. Normalize the user intent with the human-curated layer before touching the raw index.
3. Filter activities by date tags first, then primary intent, then format/location fit.
4. Promote route coherence: avoid sending the user across far venues when adjacent good options exist.
5. Return a concise ranked route:
   • 3 must-join items
   • 2 alternatives
   • one low-energy/social option
   • logistics reminder if location, pass, dining, camping, or transport matters
6. For every item include: time, location, why it matches, what this part is for, evidence source, and next action.

Tag semantics

• date_tags: exact conference date such as 2026-04-24.
• primary_topic_tags: what the activity is mainly about. Prefer human judgment from references/manual/recommendation_layer.md.
• secondary_topic_tags: useful side interests. Use raw topic_tags only as candidates.
• topic_tags: machine-generated candidate tags in activity_index.*.json; do not treat them as final when they conflict with title/summary/container.
• format_tags: participation form, e.g. forum, roundtable, workshop, exhibition-demo, meetup.
• container: 2050 board such as 新生论坛, 探索空间, 思想约会.
• crosswalk_status: article-to-activity mapping quality. If it is partial_manual or needs_manual_match, do not present it as complete.

Intent normalization

• "低强度", "轻松", "休息一下", "社交恢复", "随便逛逛": prefer life-sports, meetup, sports-camp, 热带雨林, 星空露营, 青春舞台, then avoid dense forum blocks unless the user asks.
• "能看能玩", "体验", "展台", "项目展示": prefer 探索空间 and exhibition-demo.
• "深聊", "圆桌", "观点碰撞", "哲学": prefer 思想约会 and roundtable / philosophy-mind.
• "报告", "主题论坛", "系统了解": prefer 新生论坛 and forum.

Output style

Use Chinese by default. Be concrete. Prefer schedule facts over generic summaries.

先确认：他山世界是否已注册？如果未注册，先给预览路线。

你的标签：...

推荐路线：
1. 时间｜地点｜活动名
   - 为什么适合你：
   - 这 part 是干嘛的：
   - 证据：

可选替代：
- ...

需要补查：
- ...