session-tracker

Purpose

This skill gives Bach a real-time picture of how much of his Claude Pro session budget he has consumed and how much context window memory is in use — so he can make smart decisions about when to start a new chat, what to prune, or whether he is approaching the 5-hour rolling limit.

Owner: El-Bach | GitHub: El-bach

Core Concepts You Must Know

Token Estimation Rules

Shortcut for prose estimation: word_count × 1.33 ≈ tokens

What Burns Tokens Every Turn

Every single message exchange consumes:

1. User input tokens — your prompt this turn
2. Assistant output tokens — Claude's reply this turn
3. Full conversation history re-sent — ALL prior messages, both sides, re-tokenized and sent to the model on every turn. This is the silent killer.
4. System prompt — always included, ~2,000–4,000 tokens baseline (not visible to user)
5. Project/Knowledge Base content — if active, cached after first hit, ~100–300 tokens per cached call thereafter

Context Window vs Usage Limit — These Are Different Things

Context Memory Consumption by Category

System prompt (always):          ~3,000 tokens (baseline)
Per conversation turn:
  - Short exchange (< 100 words each):   ~300–500 tokens/turn
  - Medium exchange (100–300 words):     ~800–1,500 tokens/turn
  - Long exchange (300+ words):          ~2,000–4,000 tokens/turn
  - Code-heavy turn:                     ~3,000–8,000 tokens/turn
  - File upload (10-page PDF):           ~8,000–15,000 tokens (one time)
  - Image upload:                        ~1,500 tokens (one time)

Peak Hours Impact (Beirut time, EEST)

During 3:00 PM – 9:00 PM Beirut, the 5-hour session limit burns faster than outside these hours. Anthropic throttles during peak (5am–11am PT). Budget that burns in 5 hours off-peak may burn in 2–3 hours during peak.

Execution Protocol

When this skill is triggered, you MUST:

Step 1 — Reconstruct the Session

Count every distinct message exchange (user turn + assistant turn) in the current conversation. Identify:

• Total number of turns
• Rough word count per turn (estimate from visible length)
• Any file uploads (note type and approximate size)
• Any image uploads
• Whether a Project/Knowledge Base is active

Step 2 — Estimate Token Usage

Apply the estimation rules above. Build this breakdown:

SYSTEM PROMPT (baseline):    ~3,000 tokens

TURN-BY-TURN BREAKDOWN:
Turn 1  [User: ~X words / Assistant: ~Y words]     ~Z tokens
Turn 2  [User: ~X words / Assistant: ~Y words]     ~Z tokens
...
Turn N  [User: ~X words / Assistant: ~Y words]     ~Z tokens

UPLOADS:
  [filename or type]                               ~Z tokens

SUBTOTAL (net new tokens this session):            ~Z tokens

Step 3 — Calculate Cumulative Context Load

Context load = sum of ALL turns + system prompt + uploads. This is what gets re-sent to the model every single turn.

CONTEXT WINDOW STATUS:
  Total context loaded:    ~Z tokens
  Context window limit:    200,000 tokens
  Context used:            Z%
  Context remaining:       ~Z,000 tokens (~Z more medium turns)

Step 4 — Estimate Session Budget Impact

For Claude Pro, treat the session budget as relative:

• Light session: < 30,000 tokens total → well within budget
• Medium session: 30,000–80,000 tokens → using a meaningful chunk
• Heavy session: 80,000–150,000 tokens → approaching limits, consider new chat
• Critical: 150,000+ tokens → near context limit, definitely start fresh

Map to an approximate % of a typical Pro session window.

Step 5 — Output the Dashboard

Produce a clean, scannable dashboard. Format exactly as below:

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
📊 SESSION USAGE DASHBOARD
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

SESSION OVERVIEW
  Turns completed:     N exchanges
  Uploads:             [list or "none"]
  Project active:      Yes / No

TOKEN BREAKDOWN
  System prompt:       ~3,000 tokens
  Conversation:        ~X,XXX tokens
  Uploads:             ~X,XXX tokens
  ──────────────────────────────
  TOTAL THIS SESSION:  ~XX,XXX tokens

CONTEXT WINDOW (Memory)
  Used:    XX,XXX / 200,000 tokens  (XX%)
  [████████░░░░░░░░░░░░] XX%
  Remaining: ~XX,XXX tokens
  ≈ room for ~XX more medium exchanges

SESSION BUDGET (5-hour rolling window)
  Estimated load:  [Light / Medium / Heavy / Critical]
  Budget used:     ~XX% of typical Pro session
  [████░░░░░░░░░░░░░░░░] XX%

⏰ PEAK HOURS CHECK
  Current Beirut time context: [peak = 3pm–9pm / off-peak]
  [Warning if in peak hours: budget burning faster]

💡 RECOMMENDATIONS
  [1–3 specific, actionable recommendations based on the actual state]
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Recommendation Logic

Generate recommendations based on thresholds:

Context < 30% used: → "You're well within limits. Continue normally."

Context 30–60% used: → "Moderate load. If this is a long work session, consider starting a new chat for unrelated topics to keep this thread focused."

Context 60–80% used: → "Context is getting heavy. Claude may start losing early context. Start a new chat if you need to introduce new topics. For this thread, avoid re-uploading files already shared."

Context > 80% used: → "⚠️ High context load. Start a new chat soon. Summarize key decisions from this session before closing it."

Context > 90% used: → "🚨 Critical. Claude is likely already losing early context. Start a new chat immediately. Copy any critical outputs now."

Peak hours active: → "You're in peak hours (3pm–9pm Beirut). Your 5-hour session budget is burning faster than normal. Heavy tasks are better run before 3pm or after 9pm."

Uploads present: → "File uploads stay in context for the whole session. Don't re-upload the same file — reference it by name instead."

Code-heavy session: → "Code turns are expensive. Avoid pasting entire files when only a function is relevant. Scope your context to only what Claude needs to see."

Accuracy Disclaimer

Always append this note at the bottom of the dashboard:

⚠️ These are estimates based on word count and known tokenization patterns. Actual token counts require the API. Margin of error: ±15–25%. For exact tracking, use the Claude API with token counting enabled.

Trigger Phrases

Activate this skill when the user says any of:

• "how much have I used"
• "usage so far"
• "how much context is left"
• "how many tokens"
• "session budget"
• "am I close to the limit"
• "context window status"
• "memory used"
• "كم استخدمت" / "كم باقي" (Arabic equivalents)