Fundamental skill
Check the gauge before you hand over the work.
Every AI tool runs inside limits: a context window, plan or rate limits, and sometimes a spend budget. This page teaches one durable habit — check your tool's own gauge before long work — and shows where that gauge lives in each tool Groundwork covers.
#How this failure actually feels
Three hours into a build session, the agent that was sharp all morning starts to slip. It re-suggests an approach you both abandoned an hour ago. It forgets a constraint you stated twice. Then, mid-edit, it stops and tells you to come back in two hours. Nothing crashed, and no error named the problem — the tool just got worse, then quit.
That session hit two different limits blind: the context window filled with dead ends until the plan fell out of it, and then a rate limit ended the day. Both were visible on a gauge the whole time. This page is about reading that gauge the way you'd check fuel before a road trip — a five-second habit that decides whether long work finishes.
#The three limits
| Context window | How much the model can hold right now: your conversation, the files it read, the logs it dumped. As it fills, the tool compacts or summarizes earlier material and has less room for new files and reasoning — the work degrades before anything visibly breaks. |
| Plan & rate limits | How much you can use per session, day, or week on your subscription. Hit one mid-task and the work pauses until the window resets, whether or not the diff was finished. |
| Spend budget | For API keys and metered plans: actual money per token. Mostly invisible on flat subscriptions, decisive when you wire an agent to an API key. |
#The habit: gauge first, then work
Before starting anything you expect to take more than a few minutes, check your tool's own readout of both numbers — window and plan. It changes your decisions in concrete ways: start big work in a fresh, well-briefed session instead of one that's nearly full, and don't open a multi-hour task minutes before your rate window resets.
Try it now: read your gauges
Open the tool you actually use. In Claude Code, run /context (window) then /usage (plan limits). In Codex CLI, run /status then /usage. Two numbers, five seconds. If either surprised you, this habit just paid for itself.
#Where the gauge lives, tool by tool
Every maintained tool ships a first-party gauge; learn yours. The command names will change — the habit won't.
| Claude Code | /context shows window usage, /usage shows plan limits, and /statusline can put an always-on readout at the bottom of every session. |
| Codex CLI | /status reports session config and remaining context capacity in one screen; /usage shows account usage. |
| Cursor | The usage dashboard in settings tracks plan consumption; see Editors & AI. |
| Grok Build | Same names as Claude Code: /context for window usage, /usage for credit usage and billing. |
| Local & open models | A different budget entirely: memory and speed instead of tokens and dollars. See below. |
#Reset on pollution
Wrong turns, giant log dumps, and abandoned approaches don't leave the window when you move on — they sit there and keep steering the model. That's why the agent re-suggests the dead end: it's still reading it. Clearing, compacting, or restarting with a clean summary is a skill, not a defeat.
The default move when a session fills mid-task is not to compact and limp along: start a fresh session and re-onboard it from repo state — the plan you wrote down, the commits, the notes in AGENTS.md — the same way you'd brief a teammate. Compact deliberately, with focus instructions, only when the in-session reasoning itself is what you need to keep. AI agents & rules covers this handoff pattern in depth.
Try it now: reset deliberately
Next time a session goes down a wrong path, don't just change the subject. In Claude Code, run /compact focus on the working approach, drop the failed attempts — or start a fresh session and point it at your plan file. Compare how the next answer reads.
Community tools that parse a vendor's local log files break every time the format changes. Fine as personal experiments; don't build habits — or team process — on them. The gauges above are the vendor's own and move with the product.
#Local models: a different budget
Run an open-weights model on your own machine and the token meter disappears — but the budget doesn't. It becomes your machine's memory (can the model fit?) and speed (tokens per second — can you stand the wait?). The gauge changes with it: ollama ps shows which models are loaded, their size, and where they're running, and btop shows what that's doing to the rest of your machine. Same habit, different dial. See GLM & open models for the full picture.
Why this counts as fundamentals: you are the senior partner. Knowing the machine's limits before assigning it work is the same judgment you'd apply to any teammate — token budgets are one half of that awareness, and your machine's own CPU and memory are the other.
#Related pages
- AI agents & rules — sessions, handoffs, and the repo as the memory that outlives any context window.
- Editors & AI — the tools this habit applies to, and what agents cost your machine itself.