The Walkie blog
Measured numbers on tokens, screenshots, and what your coding agent can actually read. Every figure is measured or calculated with the formula shown; sources are linked.
What is visual feedback for AI coding agents?
Show an AI coding agent what is wrong on screen instead of describing it: the four kinds of tools, what each costs, and why 3,500 tokens beats 80,730.
Open → Start hereVisual context token cost benchmark: 2026 edition
What one screenshot, four screenshots, a 30-second frame dump and a Walkie bundle cost in tokens on Claude, GPT and Gemini, from each vendor's rule.
Open → Start hereScreenshot token cost calculator
Paste your image size — get the exact token cost, by Anthropic's image-patch formula.
Open →Token Economics
What screenshots, recordings and sessions cost, in numbers you can check.
Why UI work burns more Claude Code tokens than backend work
A screenshot is not paid for once. In a multi-turn session it is re-sent every turn, so 2,691 tokens becomes 26,910 over ten. The math, and the fix.
Token EconomicsWhy your Claude Max plan runs out faster than you expect
Claude usage limits are token-based, not message-based. Long visual sessions compound fast, and the fixes cost nothing.
Token EconomicsWhat Is a Context Window? (And Why Filling It Hurts)
A context window is everything one request holds. Claude models now run 200,000 or 1M tokens, and one 30-second frame dump costs 80,730 of them.
Token EconomicsWhat does one Claude Code session actually cost?
A line-item breakdown of a Claude Code session: what is fixed, what is not, and why 2,691-token screenshots often dominate the bill.
Token EconomicsVideo, Screenshots or Words: The Real Cost to an AI
What a frame-dumped video, screenshots and spoken words each cost an AI coding agent to read, worked from Anthropic's published patch formula.
Token Economics11 Ways to Cut Claude Code Token Costs
Eleven ways to cut Claude Code token costs, ranked by tokens saved per minute of effort. Ten are free; one costs money.
Token EconomicsClaude Code token costs: where your context actually goes
A 1080p screenshot costs 2,691 tokens on Claude 4.7 and later. Where the rest of a Claude Code session's tokens actually go, line by line.
Token EconomicsWhy Your Coding Agent Can't Watch a Screen Recording
Claude accepts images, not video. What actually happens when you frame-dump a screen recording for your agent, and what the frames cost.
Token EconomicsHow Many Tokens Does a Screenshot Cost Your AI Coding Agent?
A 1080p screenshot costs 2,691 tokens on Claude 4.7 and later, capped at 4,784. The formula, a measured test, and what to send instead.
Token EconomicsHow Claude Calculates Image Tokens (the Formula, Explained)
Claude prices an image at one token per 28x28-pixel patch, capped at 4,784. The formula worked on a 1920x1080 screen and a measured 1288x811 test.
AI Coding Workflow
How to run a talk-point-draw feedback loop with an agent, and which tools fit which job.
What is visual feedback for AI coding agents?
Show an AI coding agent what is wrong on screen instead of describing it: the four kinds of tools, what each costs, and why 3,500 tokens beats 80,730.
AI Coding WorkflowAI coding agent glossary: 24 terms, defined
Token, context window, image tokens, MCP, skill, compaction, prompt caching, REVIEW.md, frame dump: 24 terms defined in under 75 words each.
AI Coding WorkflowMCP Explained (For People Who Just Want More)
MCP is an open-source standard, hosted by The Linux Foundation, that lets an AI agent call real tools and data instead of only generating text.
AI Coding WorkflowThe vibe coding workflow: a working setup for solo builders
Vibe coding is describe, generate, look, react. The look-and-react half has almost no tooling, and one screenshot costs 2,691 tokens before you speak.
AI Coding WorkflowVibe Coding Tools in 2026: The Honest Stack
The 2026 vibe-coding stack has four jobs: agents, voice input, review and testing. Every price here was re-read on the vendor page on 16 August 2026.
AI Coding WorkflowScreen Recording for Claude Code: What Works in 2026
Four ways to show Claude Code your screen, priced: raw video it cannot open, a frame dump at 80,730 tokens, screenshots, and a bundle at about 3,500.
AI Coding WorkflowCan you feed a Loom video to an AI coding agent?
Loom's AI reads the transcript, not the pixels. Frame-dumping a 30-second Loom costs 80,730 tokens; a narrated bundle costs about 3,500.
AI Coding WorkflowBest Tools for Giving AI Coding Agents Feedback (2026)
Eleven tools that turn your screen into feedback an AI coding agent can read. Every price re-read on its vendor page on 16 Aug 2026: free to $39 once.
AI Coding WorkflowGive Your AI Agent Visual Feedback Without Burning Tokens
Narrate and point instead of pasting screenshots into Claude Code or Cursor. A spoken review costs about 150 tokens; one 1080p screenshot costs 2,691.
Reviewing AI Code
Seeing, describing and verifying what the agent built.
How to run a visual review with Claude Code (step by step)
One-click setup and a 30-second loop: connect Walkie, record while you talk and point, hand the bundle to Claude Code, verify the fix.
Reviewing AI CodeWhy your AI agent can't see what it built
An agent works in text and sees a render only when a tool hands it one. What Claude Code can and cannot see in 2026, and what each look costs in tokens.
Reviewing AI CodeA 12-Point Checklist for Verifying AI-Built UI
Twelve checks for the states AI coding agents most reliably miss, why each one gets missed, and the fastest way to force it to appear right now.
Reviewing AI CodeHow to review AI-generated code you didn't write
Read a generated diff in one order: data shape, then boundaries, then logic. Five failure patterns with the exact strings to grep for, and how deep to go.
Reviewing AI CodeHow to review what your AI coding agent built
What the agent already checked, what only you can catch, and a plan-execute-verify loop that closes the gap — with the token cost of every way of looking.
Reviewing AI CodeThe plan → execute → verify loop, explained
Plan and execute are easy to automate. Verification is where quality lives — and the one step an AI coding agent can only close if you hand it a check.
Reviewing AI CodeHow to Give Feedback to Claude Code That Lands
Feedback lands when it names the location, the expectation, and the gap. Most reports name only the gap. Plus the Claude Code files that make it stick.
Reviewing AI CodeHow to describe a visual bug to an AI coding agent
The spacing looks off gives an agent nothing to fix. What to say instead: a ten-row translation table and the three bug types words alone cannot carry.
Product
What a REVIEW.md bundle is and how Walkie builds one.
Comparisons
Priced, sourced side-by-sides.