skills.henryzh.dev

skills / Tooling & Setup

dayflow-pull

Pull Henry's Dayflow activity data (what he actually worked on, day by day, card by card, with accurate hours) straight from Dayflow's local SQLite DB — read-only. Use whenever Henry asks to check his Dayflow, compute real work hours, build a weekly/daily report of what he did, verify time spent on a project, or audit his activity. Computes overlap-merged wall-clock hours (Dayflow's own sums double-count) and flags corrupt cards instead of trusting the LLM summaries at face value.

View folder on GitHub ↗ Raw SKILL.md ↗

dayflow-pull

Read Henry's Dayflow data directly from its structured store — do not eyeball the app or trust Dayflow's own hour totals. Dayflow is a screen-recording time tracker whose cards are LLM/OCR-generated, so two things must be handled:

  1. Overlapping cards — Dayflow sometimes emits duplicate/overlapping cards for the same span (esp. after re-analysis). SUM(end_ts - start_ts) double-counts them. Always interval-merge for real wall-clock hours.
  2. Corrupt timestamps — occasional cards span days (a "System" card of 1400+ hours was observed). Exclude any single card longer than ~3h and flag it.

How to use

Run the helper (read-only; safe while Dayflow is running):

python3 ~/.claude/skills/dayflow-pull/pull.py                    # last 8 days, Work, per-day merged hours + cards
python3 ~/.claude/skills/dayflow-pull/pull.py --from 2026-07-08 --to 2026-07-14
python3 ~/.claude/skills/dayflow-pull/pull.py --day 2026-07-13   # one day, card by card
python3 ~/.claude/skills/dayflow-pull/pull.py --category all     # every category, not just Work
python3 ~/.claude/skills/dayflow-pull/pull.py --obs              # also dump the rawer `observations` layer

Categories: Work (default), Personal, Distraction, Idle, System, all.

Data model (for ad-hoc queries)

DB: ~/Library/Application Support/Dayflow/chunks.sqlite — open mode=ro only.

  • timeline_cards — the summary cards: day, start_ts/end_ts (unix), title, summary, detailed_summary, category, subcategory, metadata (JSON incl. appSites), is_deleted (filter =0).
  • observations — the finer, rawer layer beneath the cards (per-batch), joined via batch_id; use to sanity-check what a card claims vs. the underlying evidence (llm_model records which model produced it).
  • llm_calls — Dayflow's own LLM invocations, if you need to audit how a card was derived (the "OCR-LLM might be wrong" check).
  • journal_entries, day_goals, daily_standup_entries — journal/goals.

Rules

  • Read-only, always. sqlite3 -readonly / ?mode=ro. Never write; the DB is live.
  • Report merged hours, not summed. State when you excluded glitch cards.
  • Read the cards, don't hand-wave. Cite actual card titles/times, and drop to observations when a card's claim looks off — that's the point of reading the structured data instead of guessing from a summary.