Your personal operating system for career transition.

Count Tokens Accurately for Prompt Budget Management

Token counting is a foundational skill in AI engineering. You need it to prevent context window overflow, to estimate request costs before sending them, and to enforce token budgets when assembling prompts with multiple components. The naive approach — counting words or characters — is wrong often enough to cause real problems.

What to build

Implement a PromptBudget class that manages a token budget for assembling multi-component prompts:

1. PromptBudget(total_tokens: int, reserve_output: int) — initialize with a total context window and an output reservation. The available budget is total_tokens - reserve_output.

2. count(text: str) -> int — use a simple approximation: len(text.split()) counts words, then multiply by 1.3 to approximate token count (tokens average ~0.77 words for English technical text). Round up.

3. fits(text: str) -> bool — return True if the text would fit in the remaining budget.

4. consume(label: str, text: str) -> bool — if the text fits, deduct its token count from the budget, record the label and count, and return True. If it does not fit, return False without modifying the budget.

5. remaining() -> int — return the number of tokens still available.

6. summary() -> dict — return a dict of {label: token_count} for all consumed components, plus a "remaining" key.

7. fit_chunks(chunks: list[str]) -> list[str] — given a list of chunks in priority order, return as many as fit in the remaining budget (greedy, no splitting).

Why this matters

Context overflow exceptions are a runtime surprise. A PromptBudget that tracks consumption at assembly time gives you the overflow check before the API call, not after. The summary() method gives you the data to log which components are eating your budget.

Constraints

• Use math.ceil for rounding up token counts.
• Do not use tiktoken or any external tokenizer — the word-based approximation is sufficient for this exercise.
• The budget should be stateful: multiple consume calls reduce the remaining budget cumulatively.