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

| Content type | Tokens per unit | |---|---| | English prose | ~1 token per 0.75 words (or ~1 token per 4 characters) | | Arabic text | ~1 token per 0.5–0.6 words (scripts tokenize larger) | | Code | ~1 token per 3–4 characters | | File upload (PDF, DOCX) | Estimate from character count ÷ 4 | | Image upload | ~1,000–2,000 tokens flat per image regardless of size |

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 Window (Memory) | Usage Limit (Budget) | |---|---|---| | What it is | How much text Claude can hold in one session | How many tokens you can consume per 5-hour window | | Claude Pro limit | 200,000 tokens total | Unpublished, ~5× Free tier | | When you hit it | Claude can't read early messages anymore | Claude stops responding, tells you to wait | | How to fix | Start a new conversation | Wait for the 5-hour window to roll | | Warning sign | Claude "forgets" earlier context | "Usage limit reached" message |

Context Memory Consumption by Category

System prompt (always):          ~3,000 tokens (baseline)
Per conversation turn:
  - Short exchange ( 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)

## Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

- **Author:** [El-Bach](https://github.com/El-Bach)
- **Source:** [El-Bach/Session-Tracker](https://github.com/El-Bach/Session-Tracker)
- **License:** MIT

Install and usage instructions live in the source repository linked above.