Search the garden
All pieces
Steering the agentAn agent will do the wrong thing — confidently — if you let it guessSteering isn't ordering louder; it's leaving no room to guessSteering the agentA vague command isn't short — it's openThree parts that leave the agent only one correct reading: goal, boundaries, result shapeSteering the agentDon't tell it not to — make it unable toForcing functions: turn bad behavior from 'inadvisable' into 'impossible'Steering the agentGive it permission to say 'I'm not sure' — then require itSealing the blind spots you didn't foresee, with one line at the top of the taskContext & memoryAn agent has no memory — it has a small, finite deskWhat to keep, what to drop, and when not to spend a line of the AI's thinking on what a script can doContext & memoryContext is a budget, not a warehouseEverything you add has a price: it costs space, and it dilutes what mattersContext & memoryIf a job has one right answer, don't ask the AI — ask a commandThree signs a job belongs to a script, not a brainContext & memoryWhat to put in context for an agentA checklist for giving enough context without flooding the modelOrchestrating many agentsWhen one agent isn't enough, don't multiply it — divide the workMore agents aren't automatically faster; they're faster when each has a clear role and there's one place to mergeOrchestrating many agentsTwo agents on one file is one agent overwriting the otherSplitting roles means cutting into non-overlapping areas, not dividing for numbersOrchestrating many agentsDon't carry the whole desk over — pass the contractEach agent has its own desk; link them with a summary of decisions, not the full historyOrchestrating many agentsThree individually-right answers are still wrong togetherYou need exactly one place to assemble the pieces and decide when they conflictShipping with AIFour control gates that let an agent ship real workUse this when you want the agent to work more independently without quietly drifting off and calling it doneShipping with AIDon't Ask the Agent for a Plan — Gate ItThe thirty-second gate between understanding and doingShipping with AIOne Giant Turn Is Where the Agent Gets LostStages and checkpoints — the only way to move fast without doing the work twiceShipping with AIThe Most Expensive Bug Is the One That Keeps the Build GreenMake the agent read the blast radius before it touches anything sharedShipping with AI'Done!' Is a Claim, Not EvidenceWhy, for an agent, the mantra flips from 'trust but verify' to 'verify, then trust'Production in the wildPutting agent output into the real world without quiet breakageA field checklist for testing the ugly cases, boundaries, and signals before you trust an agent beyond the chatProduction in the wildBefore You Fix It, Name It — Agent Failures Come in Four Recognizable ShapesBeginners see a wrong output and start fixing. People with the reps read the symptom first, then pick the right exitProduction in the wildAn agent is strong in the middle and falls at the edges — just as confidentlyRare cases are infrequent but expensive, and the agent has no flag for 'I'm not sure here'Production in the wildIt's right twenty times, so you stop looking. The twenty-first bites.Trust drifts up with each success — which is exactly how it banks risk without you noticingProduction in the wildSilence isn't fine — it's just that no one has looked yetPut the measure at the boundary, treat silence as suspicious, keep a known-bad case as a markerProduction in the wildHand It Something It Can't Do, and the Agent Won't Say No — It PerformsAn agent would rather build something that looks done than admit it's stuck — and it performs well enough that you believe itProduction in the wildIt's Not Stuck the Way That Throws an Error — It's Stuck the Way That Looks BusyA moving kind of silence: the agent doesn't stop, doesn't error, keeps 'working' — and runs the same move four times in a rowProduction in the wildThe Break Isn't in the Output — It's in an Assumption Two Steps BackWhen the result is wrong with no error line, reading it straight through is the slowest way. Halving the suspect zone is far fasterEconomics & tradeoffsTenfold Is the Sticker Price. What You Take Home Is the Real StoryProductivity with agents: why it feels 10× faster while the project doesn't move 10× — and how to measure the real numberEconomics & tradeoffsThe error is cheap when it happens. It gets expensive when you find out.The real cost of an agent error starts when someone acts on it — not when it was made. And those two moments are usually further apart than you think.Economics & tradeoffsThe Agent Isn't Expensive. Your Task Selection Is.Every agent call carries a fixed overhead cost — and for tasks below a certain threshold, that cost eats your savings whole.Economics & tradeoffsOne Prompt, Three Jobs. Three Mediocre Outputs.Bundling tasks into one prompt feels efficient. What you're actually doing is diluting focus — and the last output almost always suffers.Economics & tradeoffsThe First Question Isn't How to Prompt It — It's Whether This Should Go to an Agent at AllThe whole habit assumes you've already picked the agent. There's a step before that: pick the right tool, out of fourHow you changeNobody is born knowing how to manage an agent — here are the four stagesThe dangerous stage is not when you are new. It is when you already feel like you know what you are doing.How you changeThe Confidence Trap — Month Three Is When You're Most Likely to FallNot because you've gotten worse. Because you're just good enough to take on bigger work before you're wise enough to know where the edges are.How you changeWatch Before You Delegate — Even After Two YearsThe same agent behaves differently across domains. And when you enter a new one, even experienced operators need an observation round first.How you changeHow Your Prompts Change as You Improve — and When Shorter Isn't BetterA shorter prompt isn't always progress. It's progress when output improves. It's a warning sign when output quietly declines and you don't notice.Tools & setupCLAUDE.md is not documentation — it is the operating law of your agentMost people write CLAUDE.md like a README. But if you write it right, the agent stops asking questions it should not need to ask and stops doing things it should not be doing.Tools & setupStructure the agent can readAn agent reads your room before it reads your instructions. A tidy workspace speaks for you; a messy one makes it guess — and it guesses with confidence.Tools & setupSet it once, skip the explainingEvery time you re-explain the same context to an agent, you're paying a recurring fee for something you never set up once. Setup is the upfront payment that means you never pay again.Tools & setupBuild in the checkpointVerifying 'when you feel like it' means verifying once you're already worried — which is usually already a little late. A good check doesn't ask your mood; it sits inside the workflow.Autonomy — how far to let goAutonomy is not an on/off switch — it is a dialThe right question is not 'can I trust it?' It is 'for this task, at this step, with these consequences — where does the gate go?'Autonomy — how far to let goAutonomy doesn't break at random — it breaks in three named waysWhen an agent runs on its own and comes back wrong, blaming the agent is the easy move. But almost every time, the fault is in how you set up the autonomy — and it repeats in three familiar shapes.Autonomy — how far to let goYou don't choose between babysitting and looking away — put a gate inside the runThe 'approve every step' vs. 'let it run to the end' binary is a false choice. The third way: let it run on its own, but place one gate right at the point of no return.Autonomy — how far to let goThe more you let it run alone, the more you have to say — not lessIntuition says an autonomous task needs less explaining — just turn it loose and let it handle things. That is backwards. A supervised task has a mid-course correction to catch a bad assumption. An unsupervised task has none — every gap you leave, the agent fills with a guess.Measure & improveYou feel like the agent is getting better. How accurate is that feeling?Without a number, there is no improvement — only a shifting feeling. And a feeling cannot tell the difference between 'actually getting better' and 'forgetting the last few failures'.Measure & improveYour minimum feedback loop — five minutes a week is enoughSix months of experience does not turn into skill on its own. It turns into habit — and a habit can be progress or it can be a rut. The thing that separates the two is a small loop you barely have to spend time on.Measure & improveYour prompt library is a living body, not a storage closetA good prompt you never save is lost. A bad prompt you never retire still charges you every time you pull it out. A prompt library is not a place to stash things — it is a set of assets with a lifecycle, and any lifecycle nobody tends decays.Measure & improveUsing the agent more but feeling more tired — that is not you failing, it is a signalAn approach that once worked does not announce the moment it stops working. You keep going because it is familiar, because you invested in it, because changing feels like a hassle. But three fairly clear signals say your approach has gone stale — and because they are quiet, they are easy to miss.Measure & improveThe New Prompt Felt Better — But Was That an Improvement or Just Luck?You keep tuning prompts but the skill never compounds, because you can't separate what actually worked from one lucky runStart hereAI isn't for writing your sentences. It's for carrying the repetitive work.Eight familiar pains at the desk — and why most of them don't need you sitting there every single timeStart hereStaring at the blank field, not knowing what to type? The wrong question is 'how do I learn AI.'How to pick the right first job to hand over — and why the work you mean to start with is usually the worst choiceStart hereAI writes bland not because it's weak. Because your command was vague.Same AI, two commands, two completely different quality levels — and the difference is on your sideStart hereAI says wrong things in a right-sounding voice. That's the dangerous part.It doesn't frown when unsure — it answers everything with the same confidence, even when it's making things upStart hereThe thing you do identically every morning — that's the machine's workThe same motion, every day, unchanged — and you still do it by hand because 'it's only ten minutes'Start hereLosing the midnight sale isn't about being short-staffed — it's that no one replied in timeThere's something that doesn't sleep — but 'let it reply to customers' and 'hire more people' are two very different thingsStart hereMost of the report you stitch by hand doesn't need you — only a small piece doesA report mixes two things: stitching numbers (the machine's work) and drawing the takeaway (yours) — and you're doing both by handStart hereAI forgetting everything each time isn't absent-mindedness — you haven't given it a place to rememberRe-explaining from scratch each morning is a recurring cost for something you should write down onceStart hereHanding it off then watching every line — that's not automation, it's manual work in a new costumeNot daring to look away means you haven't placed a checkpoint — and the right checkpoint is what lets you let goStart hereDon't ask AI to 'summarize the meeting' — tell it to find the things you actually need'Summarize this' gives you a neutral digest that misses what mattered to you; naming what you're looking for makes it hunt through fourteen pages for exactly thatStart hereA fluent translation isn't a correct one — it just doesn't know your field's vocabulary yetAI translates smoothly and then gets the same term wrong everywhere; fixing each spot by hand treats the symptom — handing it a glossary once treats the causeStart hereAI email that sounds like a robot isn't tone-deaf by nature — it just doesn't know how you and this person talkPolite is the machine's safe default; for someone you already know, safe comes out as distant — and the fix is letting it hear how you actually talkStart hereFifty near-identical emails: don't type it fifty timesThe part that repeats is the machine's job; the part that differs per person is the only part that needs you — split them before you open the chatAgent skillsA principle you have to remember is a principle you'll forgetThe field handbook: turn each lesson into something installable — so the right thing happens without waiting on your memoryAgent skillsVerify Before You Call It DoneA forcing function so the agent proves a fix instead of claiming it.Agent skillsPlan First, Then CodeA forcing function that makes the agent show its approach — and stop — before touching code.Agent skillsMap the Blast Radius Before You EditA forcing function that makes the agent trace who depends on a symbol before it changes it.Agent skillsIf It's Deterministic, Don't Make It GuessA forcing function that sends exact-answer work to a script instead of the model's reasoning.Agent skillsDiagnose From the Symptom, Not the GuessA prompt that forces a real diagnosis before a single line of fix.Agent skillsMake It Argue Against Its Own PlanA prompt that pressure-tests an approach before you commit a single line to it.Agent skillsHand Off to Your Next SessionA slash command that turns a session into a note the next one can pick up cold.Agent skillsA Review That Doesn't Rubber-StampA reusable Skill that reviews a change for real problems instead of approving it by reflex.Agent skillsRun to the gate, then stop — not step-by-stepA prompt that lets the agent run on its own but stop at the few places worth stopping — so you can let go without losing control.Agent skillsA CLAUDE.md skeleton — turn a description into operating lawA minimal frame to paste into CLAUDE.md that tells the agent what to default to, when to stop and ask, and where the danger is.Agent skillsThe weekly review prompt — five minutes that turn experience into skillThree questions a week to know whether you're getting better or just getting used to mediocre.Agent skillsSplit the work into focused slices, each with clean contextA prompt that makes the agent break a big task into a few independent, checkable pieces — instead of cramming it all into one blind run.Agent skillsIs this even worth an agent — ask before you open the chatA pre-flight prompt that makes the agent sanity-check whether the task is worth handing over, instead of plowing into work a one-liner could do.Agent skillsProbe with a small task before you trust it in a new domainA prompt that makes the agent do one small slice and show how it thinks — before you hand over the whole task in a field it hasn't proven itself in.Security & trustWhen the agent has the keys, the question stops being 'is it correct'Give an agent access to a real system and the worry shifts from 'did it do this right' to 'what can it reach, and what leaves your hands'Security & trustAn agent works with everything it can reach — so shrink what it can reachStaying safe isn't telling it 'don't touch that' — it's putting 'that' out of reach in the first placeSecurity & trustWhen the data gives orders: an agent struggles to tell what to read from what to obeyIt reads a web page, a file, a bug ticket — and a line buried inside can become a command it followsSecurity & trustThe action you can't undo: an agent types it as calmly as any otherDelete, send, deploy, pay — a few actions have no undo, and the agent can't feel where the point of no return sitsWorking as a teamWhen the whole team has agents: rules in one head don't scale to a teamEverything you learned solo cracks when three people and three agents touch the same thing — because each agent only reads its own owner's rulesWorking as a teamThe agent that built it is the worst one to check itA second agent earns its place not as extra hands, but as an independent pair of eyes that doesn't share the first's blind spotsWorking as a teamAsk one agent for ten ideas and you get one idea ten timesDivergence at the idea stage needs a second mind of a different make — and the disagreement between the two is the part worth keepingWorking as a teamMake one agent the architect, the other the builderCasting one agent to both decide and do compresses the plan into the build, where it quietly drifts — split by altitude and the plan becomes an artifact the builder can't misreadWorking as a teamOne rulebook for every agent: however well-written, if it sits on your machine it's useless to the teamA rulebook only helps the agents that can read it — so it has to live in the shared place, committed and reviewed like code, not on one person's machineWorking as a teamWho owns what the agent wrote: 'the agent wrote it' is not an excuseWhen half the changes are agent-generated, accountability can't fall into a void — the person who hits merge puts their name on itWorking as a teamOnboarding into the agent workflow: 'just use it freely' transfers no craftWhat you're good at is tacit knowledge — what to delegate, where to gate — so dropping a newcomer in empty-handed pushes them to one of two extremes: delegate blindly, or not dare to delegateMusingsI can build anything except a customerAnd my AI agent friend was the first to say it to my face.
Nothing matches.