AI Dashboard — Complete Guide

Section 1

How Data Flows Into the Dashboard

Where every number comes from — end to end.

Every time you use Claude Code (in VS Code or terminal), it silently writes a log file called a transcript. These logs are the primary data source for the dashboard.

You chat
with Claude Code

→

~/.claude/projects/
*.jsonl

Log files on your PC

→

parse_transcripts.py
Reads & extracts data

→

export_data.py
Processes & builds JSON

→

metrics_data.json
Local summary file

→

sync_to_supabase.py
Pushes to cloud DB

→

dashboard.html
Shows charts

Step 1 · `~/.claude/projects/*.jsonl` — The raw log files

Claude Code automatically creates one .jsonl file per conversation. Every message is appended to that file in real-time as the conversation happens — not daily, not batched. When you send a message it is written immediately; when Claude responds it is written immediately.

One file = one session (conversation)

~/.claude/projects/ └── d--myprojects-ai-dashboard/ ├── 0b720226-a84b-452e-88ee-e7fed92ea4b1.jsonl ← Session 1 ├── 3f91c112-... .jsonl ← Session 2 └── ... ← one file per chat

Three types of lines inside each file

Line type	When it appears	Key fields
`type: "human"`	Every message you send to Claude	`sessionId`, `timestamp`, `cwd` (your project folder)
`type: "assistant"`	Every reply Claude sends back	`usage.input_tokens`, `usage.output_tokens`, `model`, `content[]` (tool calls inside)
`type: "tool_result"`	Result returned to Claude after a tool ran	Output text of the tool — not used by the dashboard

Example assistant line (simplified)

{ "type": "assistant", "sessionId": "0b720226-a84b-452e-88ee-e7fed92ea4b1", "timestamp": "2026-06-18T07:39:49.098Z", "cwd": "d:\\myprojects\\ai-dashboard", "message": { "role": "assistant", "model": "claude-sonnet-4-6", "usage": { "input_tokens": 4521, "output_tokens": 892, "cache_read_input_tokens": 1200 }, "content": [ { "type": "tool_use", "name": "Edit", "input": { "file_path": "login.html", "new_string": "..." } } ] } }

From this one line the parser extracts: 5,413 tokens consumed, 1 tool call (Edit), 1 code change to login.html.

Step 2 · `parse_transcripts.py` — The reader

Reads every .jsonl file in ~/.claude/projects/ and extracts structured metrics from the raw messages. This is the only step that touches the raw log files.

Incremental cache — fast re-runs

It never re-reads a file that hasn't changed. It stores results in .parse_state.json and checks each file's mtime (last-modified time). If unchanged → skip. This makes repeated runs fast even with 60+ sessions.

.parse_state.json (cache file — one entry per session) { "0b720226-a84b-...": { "mtime": "2026-06-24T13:40:28.303776+00:00", ← last file modified time "data": { "session_id": "0b720226-a84b-...", "project_path": "d:\\myprojects\\ai-dashboard", "start_time": "2026-06-18T07:39:45.031Z", "end_time": "2026-06-24T13:40:28.159Z", "turns": 575, "input_tokens": 2030, "output_tokens": 1213384, "model": "claude-sonnet-4-6", "tool_calls": [ { "tool_name": "Read", "timestamp": "...", "is_mcp": false }, ... ], "code_changes": [ { "file_path": "login.html", "lines_changed": 109, "language": "HTML" }, ... ], "daily": { "2026-06-18": { "turns": 263, "input_tokens": 703, "output_tokens": 459184 }, "2026-06-19": { "turns": 89, "input_tokens": 412, "output_tokens": 184200 }, ... } } } }

What it extracts per session

Field	How it is counted
turns	Count every `role: "human"` line — one per message you send
input_tokens	Sum `usage.input_tokens` from all assistant lines
output_tokens	Sum `usage.output_tokens` from all assistant lines
tool_calls	Every `content` block with `type: "tool_use"` across all assistant lines
code_changes	Tool calls where name = `Write / Edit / NotebookEdit` — lines counted from `new_string` / `content` field
daily	Tokens + turns broken down by day (`timestamp[:10]`) within the session

Output: In-memory Python dicts — no file written. Data is passed directly to export_data.py.

Step 3 · `export_data.py` — The calculator

Calls parse_transcripts.py, then aggregates the raw session data into every number the dashboard needs. Two time windows are used depending on the metric:

What it calculates	Window	How
KPI cards (sessions, lines, tool calls, avg iterations)	Last 30 days	Simple sums and averages across sessions in the 30-day window
KPI tokens & turns	All time	Summed from the weekly summary so the KPI board always matches the weekly table total
Daily data	Last 30 days	Merges the per-day breakdown from every session — if a session spans 5 days, all 5 days get contributions
Weekly summary	All time	Groups all sessions (from install date to today) into ISO weeks (`2026-W24`) — so older weeks never disappear as the 30-day window rolls forward
Tool usage	Last 30 days	Counts how many times each tool was called; top 20 returned
Language breakdown	Last 30 days	Groups `code_changes` by file extension (`.py` → Python, `.html` → HTML) and sums lines
Iterations distribution	Last 30 days	Buckets sessions by turn count: 1, 2, 3, 4, 5, 6–10, 11–20, 20+

Output: Writes metrics_data.json to the project folder. This file is overwritten on every refresh run.

Step 4 · `metrics_data.json` — The local summary file

The output of export_data.py. Contains only computed numbers — no raw conversation text. This is the file that gets uploaded to Supabase.

Top-level key	Type	Contents
`generated_at`	string	ISO timestamp of when this file was built
`days_analyzed`	number	Analysis window (default 30)
`kpis`	object	Total sessions, tokens, turns, lines generated, tool calls, avg iterations
`daily_data`	array	One object per active day: `{ day, sessions, tokens, turns, lines }`
`weekly_summary`	array	One object per ISO week: `{ week, sessions, tokens, turns, lines }`
`tool_usage`	array	Top 20 tools: `{ tool_name, cnt }` sorted by count
`language_breakdown`	array	Lines written per language: `{ language, lines }`
`iterations_dist`	array	Session count per turns bucket: `{ bucket, sessions }`
`mcp_usage`	array	MCP server call counts (if you use MCP tools)
`ai_tools_comparison`	array	Usage summary across Claude Code, ChatGPT, Copilot (reserved fields)

Real values from this team's data

{ "generated_at": "2026-07-02T11:30:46Z", "days_analyzed": 30, "daily_data": [ { "day": "2026-06-08", "sessions": 8, "tokens": 567234, "turns": 344, "lines": 9494 }, { "day": "2026-06-09", "sessions": 6, "tokens": 363764, "turns": 392, "lines": 4260 }, ... ], "weekly_summary": [ { "week": "2026-W24", "sessions": 22, "tokens": 3310997, "turns": 2193, "lines": 29712 }, { "week": "2026-W25", "sessions": 16, "tokens": 2261895, "turns": 1481, "lines": 13333 }, ... ], "tool_usage": [ { "tool_name": "Edit", "cnt": 1373 }, { "tool_name": "Read", "cnt": 1365 }, { "tool_name": "Grep", "cnt": 432 }, { "tool_name": "Bash", "cnt": 296 } ], "language_breakdown": [ { "language": "HTML", "lines": 17188 }, { "language": "Python", "lines": 15447 }, { "language": "SQL", "lines": 2945 } ], "iterations_dist": [ { "bucket": "6-10", "sessions": 16 }, { "bucket": "11-20", "sessions": 15 }, { "bucket": "20+", "sessions": 17 } ] }

Step 5 · `sync_to_supabase.py` — The uploader

Reads metrics_data.json and pushes it to the shared Supabase database so your data appears on the team dashboard for everyone to see.

What it does	Detail
Identify you	Looks up your UUID — first checks `SUPABASE_USER_ID` in `.env` (fastest); falls back to Supabase admin API lookup by your email
Ensure role	If you have no role yet, inserts `USER` into `user_roles` (safe no-op if already set)
Upsert `member_metrics`	One row per team member — if your row already exists it is updated (merged). The entire `metrics_data.json` is stored in the `data` column (JSONB)
Upsert `member_metrics_history`	Writes one snapshot per day keyed on `team_member_name + usage_date` — powers the long-term trend charts

What gets written to Supabase

{ "team_member_name": "Sunil Carpenter", "team_member_role": "Backend Developer", "team_name": "Galaxy Web Links Dev Team", "user_id": "uuid-from-auth", "data": { ...entire metrics_data.json... } }

The team dashboard reads the data column for every member and renders the leaderboard from it.

Note: Raw transcript data stays on your local machine — only metrics_data.json (computed numbers, no conversation text) is uploaded to the shared Supabase database.

Section 2

KPI Overview Cards

The five summary numbers shown at the top of dashboard.html for the selected time period.

Claude Sessions

47

Tokens Consumed

1.2M

Lines of Code Gen

3.4K

Avg Iterations

4.2

Tool Calls

620

What each card means

Card	Source field	Meaning
Claude Sessions	`kpis.total_sessions`	Total number of separate Claude Code conversations in the period. One session = one conversation from start to finish.
Tokens Consumed	`kpis.total_tokens`	Sum of all input + output tokens across every session. A token ≈ one word.
Lines of Code Gen	`kpis.lines_generated`	Total lines Claude wrote or edited using Write/Edit tools. Reflects AI contribution to your codebase.
Avg Iterations	`kpis.avg_iterations`	Average turns (messages you sent) per session. Lower = more precise prompts.
Tool Calls	`kpis.total_tool_calls`	Total number of internal operations Claude performed (Read, Edit, Bash, Grep, etc.) across all sessions.

Input vs Output tokens explained

Type	What it is	Example
Input tokens	Text sent TO Claude	Your question + every file Claude reads to answer it
Output tokens	Text Claude sends BACK	Claude's answer + all the code it writes

Why input is usually bigger: When you ask "fix the bug in App.java", Claude reads the entire App.java file before writing even one line of fix. That whole file counts as input tokens.

Section 3

Weekly Summary Table

Your Claude Code activity grouped by calendar week — shown in dashboard.html.

What the Week column means — ISO week format

The Week column shows dates in ISO 8601 week format: YYYY-Www

2026-W23 │ │ Year Week number (1–52, starts Monday)

Each week runs Monday → Sunday. Examples:

Week	Date Range
2026-W21	Mon 18 May – Sun 24 May 2026
2026-W22	Mon 25 May – Sun 31 May 2026
2026-W23	Mon 1 Jun – Sun 7 Jun 2026
2026-W24	Mon 8 Jun – Sun 14 Jun 2026

Tip: The week value is also a clickable link — click it to open the Weekly Drill-Down for that week.

What each column means

Column	Meaning
Week	ISO week — Monday to Sunday. Clickable: opens the daily drill-down modal.
Sessions	Number of separate Claude Code conversations that week.
Tokens	Total input + output tokens consumed that week.
Total Turns	Total number of messages you sent to Claude across all sessions that week.
Avg Iterations	Total Turns ÷ Sessions for that week. Lower means more efficient prompting.
Efficiency	A color-coded tag based on Avg Iterations — see table below.

Efficiency tag thresholds

Avg Iterations	Tag	Meaning
Less than 5	Efficient	Prompts are clear and precise — Claude usually gets it right in a few turns.
5 – 9	Moderate	Some back-and-forth. Room to be more specific in prompts.
10 or more	Complex	High iteration count — either a complex task or vague prompts.

Search & pagination

Use the search box above the table to filter by week (e.g. W23, 2026). Use the page buttons below to navigate if you have many weeks of data.

Section 4

Weekly Drill-Down

Click any week in the Weekly Summary table to open a detailed breakdown for that specific week.

What opens when you click a week

A modal popup appears with four stat cards and two bar charts:

Element	Shows
📅 Sessions card	Total sessions in that week
🔤 Tokens card	Total tokens consumed in that week
✍️ LOC Gen card	Lines of code generated that week
🔄 Avg Turns card	Average turns per session (color-coded green/yellow/red)
Sessions per Day chart	Bar chart — one bar per day of the week
Tokens per Day chart	Bar chart — one bar per day of the week

Avg Turns color coding in the drill-down

Avg Turns	Color
4 or fewer	Green — efficient
5 – 7	Yellow — moderate
8 or more	Red — high iteration

No daily chart? Some weeks (especially seed/historical data) have a weekly total but no day-by-day breakdown. In that case the four stat cards still appear using the weekly total, and the charts area shows "No per-day breakdown available for this week."

How to close the drill-down

Click the ✕ button in the top-right corner of the modal, or click anywhere outside the modal (on the dark backdrop).

Section 5

AI Efficiency Score — Full Calculation

A score out of 100 that measures how actively and effectively you use AI tools. Higher = you use AI more and better.

The 4 components

1. Usage Score max 30 points

Sessions per day = total_sessions ÷ days_analyzed Score = min(30, sessions_per_day × 6)

More sessions = higher score. Using Claude Code every day gets full 30 points.

2. Output Score max 25 points

Lines per day = total_lines_generated ÷ days_analyzed Score = min(25, lines_per_day ÷ 3)

More code generated by Claude each day = higher score.

3. Prompt Quality Score max 25 points

Score = max(0, 25 − (avg_iterations − 1) × 2.5) Special case: if avg_iterations = 0 (no data) → defaults to 12 points

Fewer back-and-forth messages = higher score. Clear, detailed prompts that get the answer in 1–3 turns score highest.

Avg turns per session	Prompt Quality Score
1 turn	25 points (max)
3 turns	20 points
5 turns	15 points
11 turns	0 points (minimum)

4. Tool Diversity Score max 20 points

tool_calls_per_day = total_tool_calls ÷ days_analyzed Score = min(20, tool_calls_per_day × 1.0)

More tool calls per day = higher score. 20+ tool calls/day earns the full 20 points.

Final formula

Total Score = min(100, Usage + Output + PromptQuality + ToolDiversity) = 30 + 25 + 25 + 20 = max 100 points

Real example — score of ~70 (sample member)

Input	Value
Total sessions	90
Days analyzed	30
Lines generated	1,800
Avg turns per session	4.2
Total tool calls	460

Usage = min(30, (90 ÷ 30) × 6) = min(30, 18) = 18 pts Output = min(25, (1800 ÷ 30) ÷ 3) = min(25, 20) = 20 pts Prompt = max(0, 25 − (4.2 − 1) × 2.5) = max(0, 17) = 17 pts Tools = min(20, (460 ÷ 30) × 1.0) = min(20, 15.3) = 15 pts TOTAL = 18 + 20 + 17 + 15 = 70 → Power User

Badge levels

Score	Badge	Meaning
75 – 100	AI Champion	Heavy daily AI user, efficient prompts, high tool usage
55 – 74	Power User	Regular user with good habits
35 – 54	Active Learner	Growing usage, still building habits
15 – 34	Getting Started	Occasional use, not yet daily
0 – 14	Needs Coaching	Just starting out

Section 6

Team Dashboard — Overview & Leaderboard

Aggregates data from all team members via Supabase. Open team_dashboard.html to see this.

How team data works

Each team member runs refresh_silent.ps1 (Windows) or refresh_silent.sh (Mac/Linux) on their own machine. This syncs their metrics to the shared Supabase database:

Your machine → refresh_silent.ps1 → parse_transcripts.py → export_data.py → sync_to_supabase.py → Supabase ↓ team_dashboard.html reads all members

Team Overview KPI strip

The five summary cards at the top of team_dashboard.html aggregate data across the whole team:

Active Members

8/12

Total Sessions

340

Total Tokens

9.4M

Total LOC Gen

18K

Tool Calls

4.2K

Card	Meaning
Active Members	Members who have at least 1 session in the period, out of total members in Supabase.
Total Sessions	Sum of all sessions across the entire team.
Total Tokens	Sum of all input + output tokens across the entire team.
Total LOC Generated	Sum of all AI-written lines across the entire team.
Tool Calls	Total internal operations across the entire team.

Leaderboard columns

The main ranking table sorted by AI Efficiency Score by default. Click any column header to re-sort.

Column	Meaning
#	Rank. 🥇🥈🥉 medals for top 3.
Member	Name and role. Sortable A–Z.
Sessions	Total Claude Code conversations in the period.
Tokens	Total tokens consumed (input + output).
LOC Gen	Lines of code generated by Claude.
Avg Turns	Average turns per session. Color-coded: green ≤4, yellow ≤7, red >7.
Score	AI Efficiency Score (0–100) with color bar.
Badge	Level label — AI Champion, Power User, Active Learner, Getting Started, Needs Coaching.

Search: Use the search box above the leaderboard to filter by member name or role. Results update live as you type.

Section 7

Team Weekly Summary

Shows each member's activity broken down by week, below the leaderboard in team_dashboard.html.

Columns in the team weekly table

Column	Meaning
Member	Team member name. Clickable — opens the Weekly Drill-Down modal for that member and week.
Week	ISO week badge (e.g. 2026-W24). Shows the last 4 weeks per member by default.
Sessions	Sessions that member had in that week.
Tokens	Tokens consumed by that member that week.
Turns	Total message turns across all sessions that week.
Avg Iterations	Efficiency badge: Efficient (avg <5), Moderate (avg 5–9), or the raw number if ≥10.

Search & pagination

Use the search box to filter by member name or week number (e.g. John, W23). Use the pagination buttons to navigate through all rows.

Section 8

Export Report

Both dashboards have an Export Report button in the top-right header that downloads a spreadsheet.

dashboard.html — what's in the XLSX

Sheet	Columns
KPI Summary	Metric, Value — your overall totals for the period
Daily Data	Date, Sessions, Tokens, Lines Generated, Turns — one row per day
Weekly Summary	Week, Sessions, Tokens, Total Turns, Avg Iterations — one row per week

team_dashboard.html — what's in the XLSX

Sheet	Columns
Team Leaderboard	Rank, Name, Role, Score, Badge, Sessions, Tokens, LOC Generated, Avg Iterations, Tool Calls
Daily Data	Member, Date, Sessions, Tokens, Lines, Turns — per-member per-day
Weekly Summary	Member, Week, Sessions, Tokens, Turns — per-member per-week

Section 9

How to Improve Your Score

Practical steps to increase your AI Efficiency Score.

Quick wins

Action	Impact
Use Claude Code at least once every working day	+Usage Score (up to 30 pts)
Ask Claude to write new files, not just explain code	+Output Score (LOC Gen)
Write detailed prompts — include file name, expected behavior, constraints	+Prompt Quality Score (fewer turns)
Let Claude use tools freely on complex tasks (Read, Edit, Bash, Grep)	+Tool Diversity Score (20+ tool calls/day = full 20 pts)

Good prompt vs bad prompt example

Bad — many iterations	Good — few iterations
"Fix the bug"	"Fix the NullPointerException in OrderService.java line 45. The `order.getUser()` can return null when the guest checkout flag is true. Add a null check and return an empty Optional instead."
Results in 5–10 back-and-forth turns	Usually solved in 1–2 turns

Refresh your dashboard

After using Claude Code, run the refresh script to update all metrics:

Windows: powershell -ExecutionPolicy Bypass -File refresh_silent.ps1 Mac / Linux: bash refresh_silent.sh

Or just wait — the scheduler does this automatically Mon–Fri at 18:30 if you set it up.

AI Dashboard — Complete Guide

Contents

How Data Flows Into the Dashboard

Step 1 · ~/.claude/projects/*.jsonl — The raw log files

Step 2 · parse_transcripts.py — The reader

Step 3 · export_data.py — The calculator

Step 4 · metrics_data.json — The local summary file

Step 5 · sync_to_supabase.py — The uploader

KPI Overview Cards

What each card means

Input vs Output tokens explained

Weekly Summary Table

What the Week column means — ISO week format

What each column means

Efficiency tag thresholds

Search & pagination

Weekly Drill-Down

What opens when you click a week

Avg Turns color coding in the drill-down

AI Efficiency Score — Full Calculation

The 4 components

Final formula

Real example — score of ~70 (sample member)

Badge levels

Team Dashboard — Overview & Leaderboard

How team data works

Team Overview KPI strip

Leaderboard columns

Team Weekly Summary

Columns in the team weekly table

Search & pagination

Export Report

dashboard.html — what's in the XLSX

team_dashboard.html — what's in the XLSX

How to Improve Your Score

Quick wins

Good prompt vs bad prompt example

Refresh your dashboard

Step 1 · `~/.claude/projects/*.jsonl` — The raw log files

Step 2 · `parse_transcripts.py` — The reader

Step 3 · `export_data.py` — The calculator

Step 4 · `metrics_data.json` — The local summary file

Step 5 · `sync_to_supabase.py` — The uploader