Ten of these eleven tactics cost nothing and take under a minute each: stop frame-dumping recordings, describe a bug in words before reaching for a screenshot, crop what you do capture, and clear context between unrelated tasks. The eleventh costs money: a narrated-review tool like Walkie. It only earns its place once visual review is a daily habit, not a one-off.
Updated August 2026: every figure below is either measured directly or calculated from Anthropic's published vision formula. Nothing here is a guess. They're ranked by tokens saved per minute of effort, starting with the tactic that pays off in the time it takes to read this sentence.
Why do costs run away?
Costs run away because a Claude Code session re-sends its running context on every turn. Anything expensive you add early — a raw screen recording, four screenshots instead of one — gets paid for again on every later message, not just once when you pasted it.
This is part of the wider picture of Claude Code token costs: the first thing you add to a conversation is rarely the only thing you pay for it. The clearest example is the ordinary habit of covering a page from a few angles instead of one.
- 10,764 tokens — four full-page screenshots, no narration
- 2,691 tokens — one cropped screenshot of the same bug
- 8,073 tokens saved — by sending one screenshot instead of four
That ratio — tokens saved against minutes spent choosing what to capture — is the sorting principle for the eleven tactics below. The full arithmetic behind the screenshot number is worked step by step in how many tokens does a screenshot cost an AI agent; this article is about what to actually do with it.
These are Walkie's own measured figures and Anthropic's published formula, not an exact tokenizer count. Treat the numbers as close estimates. The ratios between them are what matters for deciding where to spend your effort.
The eleven tactics
Ranked by tokens saved per minute of effort: stop frame-dumping recordings, narrate or describe bugs before screenshotting, crop and batch what you do capture, clear context between unrelated tasks, and reach for a paid bundler only once visual review is a daily habit. Ten of the eleven cost nothing.
Stop frame-dumping raw recordings into the agent. If you're recording your screen and handing the agent the video, or its extracted frames, to "watch," stop — this is the single biggest sinkhole on the list. Saves: a 30-second review frame-dumped at one frame per second runs 80,730 tokens. Describing the same thing in words removes almost all of that cost.
Say it in words before reaching for a screenshot. If the bug is namable, like "the CTA button is red instead of teal," text says it for a fraction of an image's cost. Saves: zero vision tokens instead of roughly 2,691 for a full 1080p screenshot.
Narrate instead of screenshotting. A spoken description of what's wrong, even dictated with no special tool, carries what's wrong and where in far fewer tokens than an image. Saves: a typical spoken review transcribes to roughly 100–200 tokens, against ~2,691 for one screenshot. That's about 2,500 tokens saved per instance. For the full narrate-and-point workflow, not just the arithmetic, see give your AI agent visual feedback without burning tokens.
Send one screenshot, not four. Crop to the region that actually shows the problem instead of capturing the whole page from several angles. Saves: four uncropped screenshots run 10,764 tokens; one cropped shot runs about 2,691. That's 8,073 tokens for roughly 30 seconds of extra care.
Crop before you paste. Anthropic's vision formula prices an image by its pixel dimensions, tiled into 28×28-pixel patches at one token per patch, capped at 4,784 tokens (Source: Anthropic, 2026). A tighter crop is a smaller grid of patches, full stop. Saves: scales with how much dead space you cut, up to the point the image is already under the cap.
Point at a file path instead of pasting the file. Claude Code can read a file itself; pasting the whole thing inline just duplicates what it can already fetch, and that duplicate rides along on every later turn. Saves: whatever the file's token cost would have been, paid once instead of on every subsequent message.
Grep for the function instead of dumping the directory. A targeted search returns the ten lines that matter; a full directory listing returns everything, most of which isn't the answer. Saves: the gap between ten relevant lines and an entire tree, repeated for as long as the context stays open.
Batch your feedback into one message. Five small follow-ups each re-send the context accumulated so far; one message that lists all five issues sends it once. Saves: four re-sends' worth of everything already in the conversation.
Clear context between unrelated tasks. Starting a new session, or running your agent's reset command, when you switch from one feature to an unrelated one stops old context from riding along for no reason. Saves: it's compounding, not one-time. An unmanaged thread is how a single 80,730-token frame-dump turns into roughly 40% of a 200,000-token context window gone before the actual fix even starts.
Keep the bug report to one precise sentence. Text is already the cheapest input there is; a tight sentence versus a hedging paragraph is a small, real saving with no real downside. Saves: the smallest number on this list, and also the easiest. There's no reason to skip it.
Reach for a bundler once visual review is a daily habit. A tool that narrates and selects frames automatically, like Walkie, packages the same review into one bundle instead of a raw recording. Saves: a measured Walkie bundle runs about 3,500 tokens against 80,730 for the frame-dumped equivalent: 23× fewer, roughly $0.04 instead of $0.81 at $10 per million input tokens. The free tier (7 screen recordings and 10 screenshots, lifetime, no card, no account) covers casual use; it's $39 once, not a subscription, if it becomes how you work.
Which give the biggest return?
The top three, by a wide margin: never frame-dumping a raw recording (tens of thousands of tokens per review), narrating instead of screenshotting (roughly 2,500 tokens per instance), and sending one cropped screenshot instead of four (8,073 tokens for about 30 seconds of care). Everything past that is real, but smaller.
| Tactic | Effort | What it saves |
|---|---|---|
| Don't frame-dump a recording | None — just don't | 80,730 tokens avoided per 30-second review |
| Narrate instead of screenshot | ~10 seconds | ~2,500 tokens per instance (150 vs. 2,691) |
| One screenshot, not four | ~30 seconds | 8,073 tokens per page reviewed |
| Crop before pasting | ~10 seconds | Scales with dead space removed |
| Clear context between tasks | One command | Prevents compounding toward the ~40% share of a 200k window a single frame-dump can cost |
| Point at a file path, don't paste | Habit change | A file's full token cost, avoided on every later turn |
For the full three-way breakdown of video, screenshots and words, see video, screenshots or words: the real cost of showing an AI your screen.
Which are not worth it?
A few common instincts don't move the needle. Recompressing a screenshot's file size doesn't lower its token cost, since the formula prices pixel dimensions, not bytes. Switching models to save on image-heavy work doesn't touch an image's cost either. Deleting messages mid-conversation is a real risk for a small, unreliable saving.
- Recompressing images for file size. Token cost comes from width times height in 28-pixel patches (Source: Anthropic, 2026), not JPEG quality or megabytes. A smaller file saves disk space, not tokens.
- Downscaling your display resolution. This does lower token cost in theory, since the formula reads pixel dimensions, but once an image is already near or under the cap, the saving is small, and living at a worse resolution all day is a real cost. Cropping gets you the same result without that trade.
- Switching models specifically to save on image-heavy turns. The vision formula applies the same way regardless of which model reads the image, so this doesn't change what an image costs. It only changes the quality of the response you get for that cost.
- Deleting messages mid-conversation. It can shrink context, but it's easy to delete something the agent still needed and get a worse answer back as a result. The safer version of this idea is tactic 9: clear context between tasks, not in the middle of one.
What to do first this week
Start with the three free tactics that cost nothing and take under a minute: stop frame-dumping recordings if you ever do, describe or narrate a bug before screenshotting it, and crop to one image instead of sending four. Add a fresh session per unrelated task after that.
None of this requires installing anything. Try those four for a week before touching anything else on the list — they account for most of the savings available, and none of them change how you actually work with the agent, only what you hand it. If visual review becomes enough of your week that narrating and cropping by hand starts to feel like the bottleneck itself, that's the point where a bundler stops being overkill. Walkie's free tier — 7 screen recordings and 10 screenshots, lifetime, no card required — is the honest way to find out where that line is for you before paying anything.