time-ledger – Data & Science skill for AI agents

Time Ledger

Overview

Conversational time tracking. The user reports time in plain language — *"read papers 2h, gym 1h, did a leetcode"* — and the agent parses it into structured rows (activity, minutes, date, an optional compounding tag) and writes them to the user's own Notion database through the official Notion connector. The core design is an honesty contract: anything uncertain becomes a To-confirm row and gets batch-asked — never fabricated. A log that quietly invents durations is worse than no log.

When to Use This Skill

• Use when the user reports time spent ("read papers for two hours", "hit the gym for an hour", "coded all morning")
• Use when the user says "log it", "log my time", "time ledger", or "tidy up my time ledger"
• Use when the user asks where their time went or for a time review

How It Works

Step 1: Find and verify the database

Inspect the actual connector schema and database/data-source properties. collection:// references and expanded date keys are connector-specific, not universal Notion API formats. Verify parent identity and read back the first written row; discovery does not authorize a new database or unrelated data access.

On the first write of a session, use Notion search to find the database (type database, not a page) whose title contains "time-ledger". Read its data_source_id (the collection://... UUID) and use it as the parent for create-pages / query for the rest of the session. If more than one matches, ask the user which to use. Nothing is pasted or configured.

The companion Notion template (free, linked in the source repo) ships this field schema — select values are a controlled enum, copy them exactly:

• Entry (title) — the user's words or a clear title
• Activity (select): Reading / Coding / Practice / Fitness / Investing / Meeting / Writing / Life / Other
• Minutes (number)
• Date (date) — expand to "date:Date:start": "YYYY-MM-DD"; a bare Date value fails with HTTP 400
• Status (select): To-sort / To-confirm / Done
• Compounding (select): Compounding / Consuming / Neutral
• Notes (text) — clue from the user's words / your question

Step 2: Direct report mode

Parse the report → create-pages (parent = the data source id above). Set Status=Done for what is certain; for anything uncertain set Status=To-confirm, write the specific question in Notes, and ask it once in your reply.

Parsing rules:

• Activity: books/technical material → Reading; coding/building → Coding; leetcode → Practice; gym/running → Fitness; markets/research → Investing; meetings → Meeting; docs → Writing; meals/commute/chores → Life; otherwise → Other. Unsure → To-confirm.
• Duration cues: "two hours"=120, "an hour"=60, "half an hour"=30, For vague cues such as "a while" or "all morning", leave Minutes unset and mark To-confirm; ask rather than inserting an invented number. Use exact numbers when given.
• Date: "today" = the user's local date; "yesterday" = prior day; unspecified = today. Unsure which day → ask.
• One report, multiple blocks: "two hours on X, one hour on Y" → split into rows, but confirm whether the totals are additive (a frequent ambiguity) — default additive + one To-confirm row.
• Compounding tag: only tag the obvious — Compounding = leaves a reusable asset (learning, building, writing); Consuming = forget-on-sight entertainment (only if the user volunteers it); leave ambiguous blank.

Step 3: Batch reconcile mode

When the user says "tidy up my time ledger": query rows where Status ∈ {To-sort, To-confirm} or empty → parse each → update-page the certain ones to Done; batch all remaining questions into one message; fill answers in when the user replies.

Examples

Example 1: Direct report

User:  log it: read ML system design 2h, gym 1h, did a leetcode
Agent: Logged 3 entries ✅
       · Reading · ML system design · 120min · Compounding
       · Fitness · gym · 60min
       · Practice · LeetCode · duration unset ❓ To-confirm: how many minutes?

Example 2: Batch reconcile

User:  tidy up my time ledger
Agent: 3 rows need confirmation — answering in one go works:
       1. Tue "LeetCode" — how many minutes?
       2. Wed "reading" — book or papers (Reading), or course exercises (Practice)?
       3. "all afternoon on the deck" — Wed or Thu?

Best Practices

• ✅ Batch all questions into one message; never interrogate item by item
• ✅ Give a short receipt after logging (rows written, which are To-confirm, what the questions are)
• ✅ Copy select values from the enum exactly
• ❌ Don't fabricate a duration, category, or date when unsure — To-confirm + ask
• ❌ Don't write a bare Date property; use the date:Date:start expansion
• ❌ Don't judge the user's Consuming hours; the ledger is a mirror, not a critic

Limitations

• Requires the user's own Notion workspace, the companion database (duplicate the template from the source repo), and the official Notion connector granted access to that database — connector/skill setups currently live on paid Claude tiers.
• Self-report only, by design: it does not auto-track apps or screens (an auto-tracker knows what was open, not *why* the time was spent).
• The Compounding/Consuming tags are hand rules, not a learned model.
• Select enums must match the template; if the user renames fields, the instructions above must be mirrored to match.

Security & Safety Notes

• Mutation scope: writes and updates rows only in the single user-granted Notion database, via the official Notion connector (MCP) — no shell commands, no network fetches, no credentials handled by the skill.
• On claude.ai, Notion's write tools default to *needs approval* — the first write pops an approval prompt; this is expected, not a hang.
• The honesty contract is also a safety property: when parsing confidence is low, the skill records To-confirm and asks, rather than writing invented data into the user's records.

Common Pitfalls

• Problem: Notion API returns 400 on the date field.

Solution: Expand to "date:Date:start": "YYYY-MM-DD" — a bare Date value fails.

• Problem: create-pages succeeds but the Date column is empty (known Notion MCP issue: [notion-mcp-server#121](https://github.com/makenotion/notion-mcp-server/issues/121) — expanded date fields silently dropped).

Solution: After the session's first create, read the row back; if Date is empty, fill it with update-page.

• Problem: Search returns example-row pages instead of the database.

Solution: Filter for type database when resolving "time-ledger".

• Problem: Model clock vs user timezone differ around midnight.

Solution: Treat "today" as the user's local date; ask when it could go either way.

Related Skills

• @trading-ledger - The same "parse plain language → own Notion DB → ask instead of guessing" pattern applied to trading journals (entry thesis, plan, emotion).