How TokenSavr works
A plain-English guide to estimates, calibration, and how we account for credits across different AI platforms.
What is a strategy?
A strategy is an ordered plan that tells you exactly which AI platform to use for each step of building your idea — optimized to spend the fewest credits while still getting you to a working result.
Each step includes the action to take, which platform to use (e.g. Lovable, Claude, ChatGPT), the mode (Build, Chat, etc.), an estimated cost in credits, and the exact prompt to copy-paste.
What is a credit?
We normalize every cost into a single unit: 1 credit ≈ 1 Lovable build credit ≈ ~$0.10 of API spend. This makes it possible to compare costs across very different platforms in one number.
Per-platform credit logic
Different platforms charge in different units. We convert everything to credits using these rough conversions:
| Platform | Mode | ≈ Credits per action |
|---|---|---|
| Lovable | Build | 1–3 credits per meaningful change |
| Lovable | Chat / Plan | 1 credit per message |
| Claude / ChatGPT | Chat | $0.10 of usage ≈ 1 credit |
| Cursor / Bolt | Various | Mapped to nearest equivalent |
On the results page, the Cost by platform card sums the estimated credits for every step on that platform, so you can see at a glance where your effort (and money) goes.
How AI estimates work
When you click Generate Strategy, we ask an AI to think about your idea, pick the cheapest combination of platforms and modes, and return per-step credit estimates. The AI is instructed to:
- Always answer in credits (e.g. "1 credit", "2-4 credits", "~5 credits") — never dollars or tokens.
- Make the sum of step costs land within ±20% of the headline total (a built-in sanity check).
- Prefer cheaper modes (Chat / Plan) for thinking, and Build mode only when code actually needs to change.
Calibration explained
Every time you log what a step actually cost in the tracker, we learn a tiny bit more about your personal pattern. We compute one number per user:
avgErrorPct = average of (actual / estimated − 1) across past strategies
If your actuals consistently run 30% over the AI's guesses, the next time you generate a strategy we quietly tell the AI: "This user historically spends 30% more than estimated — bump your numbers up accordingly."
You'll see a small Calibrating estimates toast when this kicks in. It only activates when:
- You have at least 2 past strategies with logged actuals
- Your average error is at least ±10%
- You haven't disabled calibration in Settings
Want raw, unmodified AI estimates? Go to Settings and switch off Use historical calibration.
Tracking actuals
On the results page (and the dashboard), each step has a checkbox and a small "actual cost" field. Logging real numbers does two things:
- Powers the Estimate accuracy widget on your dashboard.
- Feeds the calibration loop above, so future strategies get more accurate over time.
Privacy & data
Your strategies, progress, and actuals are stored in your account and protected by row-level security — only you can read or modify them. The calibration signal never leaves your account; it's recomputed client-side and passed to the AI only as a single percentage number.
Tips for the most accurate estimates
- Be specific in the idea field — features, integrations, target users. Vague prompts get vague (and often optimistic) estimates.
- Pick the platforms you actually have access to. The AI won't route work to platforms you didn't tick.
- Log actuals as you go. Even rough numbers improve calibration after just 2-3 strategies.
- Treat the plan as a starting point, not a contract. Real software always involves a few unplanned iterations.