life-capture（生活捕捉）

life-capture

Turn natural-language life logs into durable records. This skill classifies each input item, generates tags, creates user-visible markdown, writes to a daily note under life/daily, and syncs structured data into life/db/life.db.

Default storage layout

Use these paths unless the user explicitly overrides them:

life/
  daily/
  ideas/
  db/life.db

Create missing directories as needed. Never delete existing content. Append or update only.

Supported record types

Map every parsed item to exactly one primary type:

• expense: spending, bills, purchases, subscriptions, refunds
• task: completed tasks, ongoing work, todos, chores, habits
• schedule: calendar items, appointments, time blocks, plans
• idea: ideas, inspiration, possible projects, reflections worth saving

When a sentence contains multiple items, split it into multiple records.

Output contract

For each user request:

1. Parse the message into one or more records.
2. Generate a stable id for each record using the pattern:

- exp_YYYYMMDD_NNN - task_YYYYMMDD_NNN - sched_YYYYMMDD_NNN - idea_YYYYMMDD_NNN

1. Generate 1 to 4 short tags.
2. Show the user the organized result in markdown.
3. Save the records by running scripts/process_entry.py.

Always keep the original user wording in raw_text. Never invent missing fields. Leave unknown fields null.

User-visible response format

Because this skill is configured for visible output, show a concise but complete result after writing:

## 已整理记录

### 1) <type label>
- ID: <id>
- 标签: #a #b
- 归档: <daily markdown path>
- 数据库: <written/skipped>

#### Markdown
<the markdown block written for this item>

#### JSON

<the parsed record json>

If there are multiple records, repeat the block for each one.

Parsing rules

Use scripts/parse_entries.py for natural-language parsing. The parser now reads configurable rules from references/parser_config.json, so prefer editing that file instead of changing Python when you need new categories, tags, or keyword mappings.

Expense

Extract when present:

• amount
• currency (default CNY only when the currency symbol or language implies RMB; otherwise null)
• category
• subcategory
• merchant
• pay_method

Default top-level tags often include 开销 plus one semantic tag such as 餐饮 or 交通.

Preferred categories:

• 饮食
• 交通
• 购物
• 居家
• 社交
• 娱乐
• 医疗
• 学习
• 其他

Task

Extract when present:

• status (todo, doing, done, cancelled)
• priority (low, normal, high)
• project
• due_date
• completed_at

If the user says they already did something, default status to done.

Schedule

Extract when present:

• schedule_date
• start_time
• end_time
• location
• status (planned, done, skipped)

If the user uses relative dates, resolve them from the current conversation date. Prefer passing --today YYYY-MM-DD to scripts/process_entry.py or scripts/parse_entries.py so relative dates like 明天 are stable across environments.

Idea

Extract when present:

• idea_type
• status (captured, reviewing, used, archived)
• related_task_id

Default status to captured.

Configurable parsing rules

Before editing Python, check whether the change can be made in references/parser_config.json.

You can change:

• category and subcategory mappings for expenses
• task project mappings
• idea type mappings
• schedule extra tag mappings
• default tags by record type
• hint regexes used in type inference

To test a modified config without changing the bundled default file:

python scripts/parse_entries.py --config /path/to/custom_config.json --text "买咖啡 18 元，明天下午两点去体检"

Markdown writing rules

Write each record into the daily note for its effective date under one of these sections:

• ## 开销
• ## 任务
• ## 日程
• ## 灵感

Use this block structure:

### <id>
- 时间：<time or empty>
- 标签：#tag1 #tag2
- 原始描述：<raw_text>
- 摘要：<summary>

Then add type-specific fields:

• Expense: 金额 / 币种 / 分类 / 子分类 / 商家 / 支付方式
• Task: 状态 / 优先级 / 项目 / 截止日期 / 完成时间
• Schedule: 日期 / 开始时间 / 结束时间 / 地点 / 状态
• Idea: 类型 / 状态 / 关联任务

Execution workflow

End-to-end one-command flow

Use this when the user provides natural language and wants the records saved immediately:

python scripts/process_entry.py --root life --db life/db/life.db --today 2026-03-10 --text "今天中午牛肉面 26 元，下午整理了书桌，想到可以做一个生活数据看板"

The wrapper script will:

1. initialize the database if missing
2. parse text into {"records": [...]} with scripts/parse_entries.py
3. save markdown and sqlite rows with scripts/save_entry.py
4. print parsed records plus save results as json

Split-step flow

Use this when the user asks to inspect or verify the structured output before writing:

python scripts/parse_entries.py --text "明天下午两点去体检，买咖啡 18 元"

Then save:

python scripts/save_entry.py --root life --db life/db/life.db --stdin-json

Database init only

Use this once before first write if life/db/life.db does not exist and you are not using process_entry.py:

python scripts/init_db.py --db life/db/life.db

Database sync rules

The database design is:

• entries
• expenses
• tasks
• schedules
• ideas
• tags
• entry_tags

See references/schema.md for the schema, references/examples.md for sample payloads and commands, and references/configuration.md plus references/parser_config.json for configurable parsing rules.

Failure handling

• If markdown write succeeds but database sync fails, say so clearly.
• Do not silently drop a record.
• If parsing is ambiguous, make the narrowest safe interpretation and preserve the original text.
• If a record is missing a critical type-specific field, still save the record with null fields rather than discarding it.