The words below come up constantly in Claude Code sessions, in tooling docs, and in arguments about what a session costs. Each entry is self-contained, so it can be read on its own, and carries a number wherever a real one exists. Terms run alphabetically. The wider argument they sit inside is visual feedback for AI coding agents.
Every number used more than once in this glossary:
| Figure | Value | Basis |
|---|---|---|
| Image patch formula | ceiling(width / 28) x ceiling(height / 28) | Anthropic's vision documentation |
| One 1920x1080 screenshot, high-resolution tier | 2,691 tokens | calculated |
| Per-image cap, high-resolution tier (Claude 4.7 and later) | 4,784 tokens | Anthropic's vision documentation |
| The same screenshot on standard tier, downscaled to 1456x819 | 1,560 tokens | calculated |
| A complete Walkie bundle | about 3,500 tokens | measured, typical |
| A 30-second review frame-dumped at 1 fps | 80,730 tokens | calculated |
| Dollar conversion used throughout | tokens / 1,000,000 x $10 per million input tokens | calculated |
Price any resolution against the same formula in the screenshot token cost calculator, or read the cross-vendor version in the visual context token cost benchmark.
What is an agent (coding agent)?#
A coding agent is a program that wraps a language model in a loop with tools, so it can read files, run commands and edit code rather than only produce text. Claude Code, Cursor and Codex are coding agents. The model supplies judgment; the harness supplies file access, a terminal, and permission to act. Take the tools away and you are back to a chat window that can only describe a fix.
What is an agentic loop?#
The cycle a coding agent repeats until a task is finished: read context, choose a tool, run it, read the result, decide again. Each pass is a fresh API request carrying the whole conversation so far, which is why a long loop costs far more than its final message suggests. Loops end on success, on a usage limit, or on a person interrupting them — and an unattended loop that has gone wrong keeps spending until one of those three happens.
What is a bundle (REVIEW.md bundle)?#
The folder Walkie writes after a recording. It contains REVIEW.md — the narration transcript with pointed frames referenced inline, plus a frame index table — alongside the frame images and a savings note. A complete bundle averages about 3,500 tokens (measured, typical). It is plain markdown and PNG, so any agent that can read a folder can read it, with no plugin and no proprietary viewer involved.
See: what a REVIEW.md bundle actually is
What is Claude Code?#
Anthropic's command-line coding agent. It runs in a terminal, reads and edits files in your project, runs commands, and extends through MCP servers for tools and skills for instructions. It sends the full conversation with every request, which is the single most important thing to know about its cost. Anthropic's own cost documentation puts the average across enterprise deployments at around $13 per developer per active day.
What is compaction (context compaction)?#
Summarizing older conversation history so a session can continue past the context window. Claude Code auto-compacts as the window fills, and /compact triggers it manually. Anthropic's cost documentation notes that compacting a large context is itself a large request, because /compact reads the conversation it summarizes. /clear throws history away instead, which costs nothing but loses everything the summary would have kept.
What is a context window?#
The total span a model can reference in one request: system prompt, every prior message, tool definitions, images, and the response being generated. Claude Sonnet 4.5 has 200,000 tokens; several newer Claude models have 1,000,000, per Anthropic's context-window documentation. Cached content still occupies the window — caching changes the price of those tokens, not whether they count.
What is a frame dump?#
Extracting still frames from a video at a fixed interval, commonly one per second, so an image-only model can see snapshots of what happened. A 30-second recording becomes 30 images. At 2,691 tokens per 1080p frame on Claude 4.7 and later, the high-resolution tier, where no image exceeds 4,784, that is 80,730 tokens (calculated) — and a fixed-interval sampler has no idea which of those seconds mattered.
See: why your coding agent can't watch a screen recording
What is hallucination?#
A model stating something false with exactly the fluency it uses for something true: an invented function name, a file that does not exist, a confident account of a screen it never saw. Anthropic's vision documentation lists it as a known limitation for low-quality, rotated or very small images. The defense is verification against a stated expectation, not better prompting, because a fluent wrong answer reads the same as a right one.
What are image tokens (the patch formula)?#
The unit Claude bills images in. Claude views an image in 28x28-pixel patches, so it costs the ceiling of width over 28, times the ceiling of height over 28, in visual tokens (Anthropic's vision documentation). On the high-resolution tier, Claude 4.7 and later, the long edge caps at 2,576 px and the total at 4,784 tokens; a 1920x1080 screenshot works out to 69 x 39 = 2,691.
What are input and output tokens?#
Input tokens are everything you send: system prompt, conversation history, tool definitions, images. Output tokens are what the model generates, and extended thinking is billed as output per Anthropic's context-window documentation. In an agent session the input side usually dominates, because the entire history is resent with every request while each individual response stays small. Both count against the same context window.
What are intent-selected frames?#
Frames that earn a place because a person's voice and cursor marked the moment, not because a timer fired. In Walkie, dwelling the pointer near what you are describing marks that moment, and Cmd+Shift+M marks one explicitly. The result is a handful of frames tied to what was actually said, instead of one frame per second regardless of whether anything on screen changed.
What is MCP (Model Context Protocol)?#
An open standard for connecting AI applications to external tools and data, published at modelcontextprotocol.io with reference implementations in the modelcontextprotocol GitHub organization. Anthropic's Claude Code documentation describes it as an open source standard for AI-tool integrations. It defines how a client discovers a server's tools, calls them and receives results, so one server works across every MCP-capable client.
What is an MCP server?#
A process that exposes a set of tools over MCP for an agent to call. Anthropic's Claude Code MCP documentation stores user-scoped servers in ~/.claude.json, making them available across every project on the machine while staying private to your account. Walkie ships one exposing seven walkie_* tools for recordings and screenshots; Claude Code reads its MCP configuration when a session starts.
What is the plan → execute → verify loop?#
A three-step discipline for working with a coding agent: agree what will change before any code moves, let the agent make the change, then check the result against the expectation you named. Claude Code supports the first step with plan mode. The third step is the one people skip, because watching an agent say it is done feels like confirmation. It is not.
See: the plan, execute, verify loop
What is a pointer trail?#
The path a cursor takes during a screen review, sampled while the person is talking. It is the signal that says which part of the screen was meant, which a static screenshot cannot carry — a screenshot shows twelve identical rows, the pointer says which one. In Walkie, dwelling near an element promotes that moment to a captured frame, so the trail decides what the agent ends up looking at.
What is prompt caching?#
Reusing an already-processed prefix of a request instead of paying full price to process it again. Anthropic's prompt-caching documentation prices cache reads at 0.1 times the base input rate, with a five-minute lifetime by default and a one-hour option, and lists images and documents as cacheable. Cached content still occupies the context window: caching changes the price, not the count.
What is a rate limit or usage limit?#
The ceiling that stops a session. On a subscription plan Claude Code reports it as hitting your session limit or your weekly limit, drawn from a seat allowance that resets on a rolling five-hour window and a weekly window. Switching models does not restore access, because those windows are shared across models. On API and cloud plans the equivalent controls are workspace spend limits.
What is screenshot token cost?#
What an agent pays to look at one screenshot. On Claude 4.7 and later, the high-resolution tier, a 1920x1080 screenshot costs 2,691 input tokens and no single image exceeds 4,784. Standard-tier models downscale the same image to 1456x819 and pay 1,560. At $10 per million input tokens, 2,691 tokens is roughly $0.03 — cheap once, and re-sent on every later turn.
See: how many tokens a screenshot costs
What is a skill (Claude Code skill)?#
A folder holding a SKILL.md file of instructions the agent loads when they are relevant. Anthropic's skills documentation places personal skills at ~/.claude/skills/<name>/SKILL.md, applying across all your projects, and says the frontmatter description is what tells Claude when to load one. Walkie installs walkie-review, which teaches an agent to read a bundle without opening every frame.
What is a system prompt?#
The instruction block sent ahead of the conversation on every request, setting the agent's role, its rules, and the context it starts from. It counts toward the context window like any other input and is resent each turn, which is why prompt caching targets it first. In Claude Code, your CLAUDE.md is loaded into context at session start and carries that same per-turn weight.
What is a token?#
The unit a model reads and bills in: roughly a word fragment for text, and a 28x28-pixel patch for an image. Everything in a request is counted in tokens — system prompt, every prior message, tool definitions, images, and the response being generated. Anthropic's token-counting API returns the exact count for a request before you send it.
What is tool use?#
A model choosing to call a function that was defined for it, receiving the result, and continuing from there. It is the mechanism under file reads, shell commands and MCP calls in a coding agent. Every tool result becomes part of the conversation, so a session that runs many tools grows its own input cost with each call — the results are carried forward whether or not they turned out to matter.
What is a transcript?#
The text of what was said during a screen review, produced by speech recognition. It is the cheapest carrier of intent an agent can read: a real 15-second narration measured 65 tokens, and a typical one runs about 150 (measured, range 100 to 200). In a Walkie bundle the transcript is the primary signal and the frames are supporting evidence, which is what keeps a whole review near 3,500 tokens.
What is vibe coding?#
Building software by describing what you want and reacting to what the agent produces, rather than writing most of the code yourself. The work shifts from typing to specifying and checking. That makes the review step — seeing what actually got built, and saying precisely what is wrong with it — the part that decides whether the output is any good, which is where most of the difficulty moved.
