The words that get coding agents to do what you mean. Each one comes with a prompt swap: what people usually type, and what to say instead.
What the agent can see, and how long it remembers it.
noun
Everything the model can see at once: your messages, the files it read, tool output. Past the limit, older parts get dropped or summarized.
Instead of: Remember what we changed earlier?
Context window0k / 200k
NOTES.md in memory · decision: use Pagefind
Illustrative numbers.
noun
When a session gets long, the harness summarizes older turns to free up space. Details you didn't write down can get lost.
Instead of: Keep going.
noun
Pointing the agent at a real source, like a file, a doc page or an error, instead of letting it answer from memory.
Instead of: How does the Astro content API work?
The tool around the model, and the pieces you plug into it.
noun
The program that runs the model in a loop with tools: Claude Code, Codex, Cursor. Same model, different harness, different results.
Instead of: Can the AI run my tests?
noun
A folder with instructions (a SKILL.md) that the agent loads only when the task matches its description. It costs almost nothing until it's needed.
Instead of: Always remember to check for existing helpers.
noun
A saved prompt you run by name, like /pr-description. Same words every time, no retyping.
Instead of: Write a PR description like you did last time.
noun
A script the harness runs on an event, like after a file edit or when the agent stops. A rule the model can't forget, because the model isn't the one enforcing it.
Instead of: Please always run the linter.
noun
A separate agent with its own fresh context, sent to do one piece of work and report back. Keeps the main conversation small.
Instead of: Look through the whole codebase and tell me everything.
Main conversation18k tokens
Subagent, fresh context0k tokens
Illustrative numbers.
How work gets split, handed off and checked.
noun
The note you give another agent: the goal, the files, what done looks like, and when to stop. The brief decides the result more than the model does.
Instead of: Fix the search thing.
noun
A quick throwaway build to learn one thing. No tests, no polish, deleted afterwards.
Instead of: Build the feature.
noun
Why a bug happens, as opposed to where it shows up. Fix the cause and every place it shows up is fixed too.
Instead of: Make the error go away.
noun
A code review by a different model from the one that wrote the code. The builder shouldn't mark its own work.
Instead of: Is this secure?
Shipped. Nobody else looked.
GPT Sol 6.1max
Found: usage credits can be abused by switching workspaces.
Fixed. The abuse steps no longer work.
Illustrative numbers.
Words that make a request precise.
noun
The point where the agent should stop and ask you instead of pushing on.
Instead of: Do whatever it takes.
fix the date bug in formatDate
0 tool calls
Illustrative numbers.
Instead of: Make it better.
Where the tokens go.
noun
Everything you send the model: messages, files, tool results. On long sessions this is usually most of the bill.
Instead of: Read the whole repo.
noun
What the model writes back. There are fewer of them, but each one costs more than an input token.
Instead of: Explain everything in detail.
noun
Input the provider saw recently and charges much less for. Prompts and files that stay the same hit the cache. Reworded ones don't.
Instead of: Here's the project context again, written differently.
Ready
cache read, cheapfresh inputbilled $0.00
Illustrative numbers.