Decision Pulse: AI Cognitive Load Dashboard

AI adds a lot of small decisions to a day. The agent finishes a step and stops to ask you. You keep several conversations open and hop between them. You sit for hours. None of it is hard on its own, but it adds up, and you rarely notice when you're tired.

Decision Pulse reads one day of your Claude Code and Codex transcripts, locally on your own computer, and turns it into a one-page dashboard that shows where the day got heavy.

How it reads your day

It follows a mechanism chain where each stage pushes pressure into the next:

1. Decision density: how many things needed your call (messages per active hour)
2. Recovery window: whether you actually took a break (longest stretch without a 10-minute break)
3. Cognitive fatigue: how fragmented your attention was (conversation switches per active hour)
4. Judgement quality: whether you still read carefully (share of long AI replies you answered within seconds)

Each stage gets its own low / medium / high level. They are never summed into a single score, so you can see which stage to deal with first.

What you get

• A one-line summary you can read without opening the page
• A "right now" panel when you check today: how long the current stretch has lasted
• Four colour-coded stage cards with values and the thresholds behind them
• A few suggestions for stages at medium or high, written like a colleague's reminder and each tied to a cognitive-science mechanism and paper
• Details: conversations, characters typed and read, tokens, high-stakes moments, morning/afternoon/evening split, work and break timeline, switching, prompt length and review behaviour
• Role presets (engineering, ops/HR, sales, design, manager, general) that adjust what counts as normal density, while break, switching and judgement thresholds stay the same for everyone
• An optional desktop clock (Tkinter) with four arcs showing the stage levels, refreshed every 30 minutes, with no notifications
• Full English and Traditional Chinese interface

Research basis

Thresholds cite Albulescu et al. 2022 (breaks), Blain et al. 2016 (sustained control), Leroy 2009 and Leroy & Glomb 2018 (attention residue), Mark et al. 2008 (interruptions), Parasuraman & Manzey 2010 (automation complacency) and others, each with a stated confidence level. Thresholds without research behind them are labelled "rule of thumb". Ego depletion, the 90-minute ultradian rhythm and Pomodoro are deliberately not used.

Privacy

• Reads only local transcripts in ~/.claude/projects and ~/.codex/sessions
• The script makes no network calls and uploads nothing
• Daily results stay in the skill's own history/ folder on your machine

What it doesn't do

It doesn't judge whether your decisions were right, compare days, push reminders, give a single overall score or guess your emotions. Suggestions are fixed texts triggered by rules, not generated on the fly.

Created and maintained by Wesley Cheng, author of github.com/chengwesley/decision-pulse (MIT).

繁體中文說明

Decision Pulse 讀取你電腦上當天的 Claude Code 與 Codex 對話紀錄，依「決策密度 → 恢復窗口 → 認知疲勞 → 判斷品質」這條機制鏈產出一頁儀表板。四個階段各自分成低／中／高，不加總成單一分數；中或高的階段會給出有認知科學依據、像同事提醒般的建議。全程在本機執行、不連網、不上傳。介面完整支援繁體中文與英文，也可依職能（工程、營運・HR、業務、設計、主管）調整決策密度門檻。