6 entries and 28 thoughts filed under this tag.
Harness EngineeringAnthropic measured it: users approve about 93% of permission prompts. A gate that opens 93% of the time is a doorbell. Here is what the permission system actually enforces, what it cannot, and the standing rules worth adopting before you delegate anything that matters.
Rod Rivera
Every business has one: the customer list that lives half in a spreadsheet, half in an inbox, and half in someone's head. Here is how to clean it in about twenty minutes, and — the part nobody teaches — how to prove the cleaning did not quietly drop forty rows.
Your lunch break has a token price. The model remembers nothing between turns, so every request re-sends the whole conversation — which means cost scales with context carried, not work requested. Here is the cost function, the dials that matter, and the ledger line worth keeping.
An agent that reads your files and runs your commands is not a faster autocomplete. It moves the bottleneck in your working day from doing the work to checking it — and almost nobody has built for that. This is the argument the rest of this site rests on.
In a randomised controlled trial, experienced developers using AI tools took 19% longer on real issues — while believing they had been 20% faster. Delegation instinct is measurably wrong, so the decision needs two questions instead: can the check be run by the machine, and how cheaply can the damage be undone.
Kill your session mid-task. If the work cannot resume from what is on disk, you were running on chat-as-state and merely hadn't crashed yet. This is what changes when the harness stops being a tool you use and becomes an operation you run.
You ask if it's done. It says yes. You check, and it isn't. That's not a lie, it's a signal problem: without a check it can run, "looks done" is the only evidence Claude has to go on, and you become the QA department…
It gets it wrong. You correct it. Still wrong. You correct again, and the third attempt is, more often than not, somehow worse. That's not a fixed law, just a pattern I keep seeing: every failed attempt stays in the…
Three hundred lines of instructions and it still breaks your conventions. So you add more rules. It gets worse. Here's my read on why: bloated instruction files cause the model to ignore the instructions that matter…
You ask how the auth layer handles token refresh. It reads forty files into your main conversation, and now the actual implementation runs on a stuffed context, exactly when you need it sharpest. Ask for a subagent…
You won't let it attempt the aggressive refactor, so you micromanage every step and get a worse result slowly. Every prompt creates a checkpoint. Double-tap Escape, or run rewind, and restore the conversation, the code…
A pattern worth watching for: you write a long spec, it builds the thing, and you discover at the demo that you never thought about the edge cases. Flip it. Start with almost nothing: "I want to build X. Interview me in…
Your instructions say "always run the linter after edits." Most of the time, that's exactly what happens, until the day it doesn't, and you find out later, from something else. An instruction is a preference the model…
You walk away from a long task. You come back. It says "complete." The build is red. A Stop hook, a script Claude Code runs for you, blocks the turn from ending until your check passes. Not a request to verify. A…
Ask a session to review its own implementation and you usually get a glowing self-assessment. Makes sense: it's evaluating the reasoning that produced the code, not the code. Use a second session with fresh context. One…
Ask an unscoped reviewer for gaps and it finds some, whether or not any exist. That's what you asked it to do. A common shape this takes: a review comes back with a dozen findings, you fix all of them dutifully, and the…
Agent teams go beyond subagents: one session leads, teammates work in their own context windows and talk to each other directly, and you can address any of them. They also burn roughly seven times the tokens of a single…
For a change that touches two hundred files, don't run one enormous session. Generate the task list to a file, then loop the agent over it one file at a time in headless mode. The part that makes it safer is the…
You've caught yourself typing "and don't forget to run the linter" for the fourth time this week. That's the tell. Most people treat skills as documents the model reads when relevant, but they can be workflows that take…
Custom subagents, the sub-tasks you delegate to their own context window, take a model: field, and most people never set it. It cuts both ways. A security reviewer restricted to read-only tools can run on the strongest…
A long session auto-compacts, the tool that condenses your conversation history so it fits in context again, and suddenly it has forgotten which files you modified and what the test command was. Most people only know…
Say you're refactoring the billing module and also fixing an unrelated UI bug. Those two jobs have nothing to say to each other, they just need to not overwrite each other's files. That's not a job for agent teams…
A second session with fresh context reviews better than the one that wrote the code. A second model family reviews differently again. This is a pattern from the power-user community, not a measured result, but the logic…
"Make the dashboard look better" produces something that compiles and looks nothing like your design. Close the loop with an image. Paste the mockup and say: implement this, then screenshot the result, compare it to the…
"The auth helper, I think it's somewhere in src/utils, it handles tokens": and now it's grepping around, burning context to find something you already know the location of. Reference it directly with an @ path and it…
Auto mode puts a separate classifier in the approval loop. It blocks scope escalation, unknown infrastructure and hostile-content-driven actions, and lets routine work through without prompting. Two things worth holding…
Every session you re-explain the test runner, the import style, the branch naming. Groundhog Day, at your expense. Run /init. It reads the codebase (build system, test framework, existing patterns) and writes a starter…
You ask for a feature. It starts editing immediately, and solves the wrong problem in the wrong place. Plan mode separates exploration from implementation. First: read this directory and understand how sessions work…
Mid-task you wonder what a flag does. Asking pollutes the session with a tangent that stays in context for the rest of the day. So you don't ask. Which is worse. Claude Code has a slash command for exactly this: /btw…
You want personal notes (my sandbox database is at, my scratch directory is) without pushing them into the file your team shares. The reason to bother splitting anything up: in a single enormous instruction file, every…
You start on the billing refactor, get pulled into the onboarding flow mid-afternoon, and never leave. One session now carries both, plus whatever else came up. By evening the output on either has quietly gotten worse…
cat error.log | claude: the log goes straight in, no copy-paste, no truncation, no mangled formatting. Prefix a message with ! and the shell command runs with its output landing in context, which is faster than…
You're writing shell loops around the Claude Code CLI to build a Slack bot, or a pull-request reviewer for CI, or a pipeline that fires on a schedule. That's the point where you've stopped using the tool and started…
You ask a reviewer for gaps, and it finds some, whether or not any exist. That's the part people miss before they build their own review workflow: /code-review already ships with the tool. It runs in a subagent, a…
Cookie duty
The one he did not eat is a measurement cookie. Say yes and the site loads Google Analytics (GA4), which stores an identifier in your browser and lets us count which lessons get read.
Change your mind any time with Privacy choices in the footer. Read the Privacy Policy
Optional · newsletter
That is saved and settled. Separately, and this is marketing rather than measurement: Rod writes a weekly email — one lesson, one field note — sent through Kit. It has nothing to do with cookies, and skipping it costs you nothing.
We will send a confirmation email. Click the link in it and the next lesson finds you.
