AI Agent Hacks That Work (and the Ones to Skip): A Sorted, Checked-Against-the-Docs Guide

Which AI agent hacks hold up (the sorted list first)

A first agent usually goes wrong in a predictable way. It gets a loose instruction, nothing to check itself against and a conversation that keeps growing, and what comes back looks finished without being right. Most popular agent habits are repairs for that one pattern.

Sorted by the job they do, the habits fall into three groups: ones that say what good looks like, ones that check the work, and ones that change what the work costs. We read Anthropic’s documentation, the tool makers’ pages and, where one exists, the original paper for each habit on October 19, 2026. We haven’t run a controlled test, and nothing here is our benchmark.

“Works” means the tool’s own documentation recommends it. “Caveat” means it carries a cost the hype leaves out. “Skip” means we found no support for the usual claim.

Nine came through clean. Each says what good looks like, adds a check, or controls what loads into the agent’s context. The four caveats cluster around power and cost, where the claims are biggest, and the one skip is the persona.

Hacks for defining the job: done, not done and a shared vocabulary

A finance analyst asks an agent to tidy the variance table and gets back a different table with better headings. The instruction had no finish line, so “done” meant whatever the agent decided.

The first habit is to write down both ends before the agent starts: what a finished result looks like and what a miss looks like, with an example of each. Anthropic’s best-practices page makes the same point for coding. Without a check the agent can run, “looks done” is the only signal available, and you become the checking step.^[1]

You’ll also see this wrapped in threats and rewards, such as telling the agent it’ll be shut down if it fails. A prompt changes what the model sees on one request. It doesn’t retrain the model, so there’s no learning loop for a reward to feed. The best-known evidence is a 2023 paper that added emotional phrases and reported gains on 45 tasks with models of that time, including ChatGPT and GPT-4.^[2] Those phrases were encouraging, none of them threats, the authors call it uncertain whether models grasp emotion, and a plain finish line does the same work.

The second habit is a glossary for words that mean different things to different people, such as “recent”, “at risk” or “good”. Let the agent ask you what you mean, then store the answers in your instructions file. Anthropic describes that file as context, not enforced configuration, so a glossary guides the agent without guaranteeing anything.^[3] Our explainer on CLAUDE.md and MEMORY.md files covers what belongs in it.

Hacks for shaping the agent: an interview, a sketch and a skill that builds skills

The second group shapes the agent itself, and each habit makes the person do the thinking before the agent does the typing.

The first is to let the AI interview you instead of writing the brief from a blank page. Anthropic recommends this for larger tasks: start with a short prompt, let the agent ask about the hard parts you hadn’t considered, and have it write the spec before building. In Claude Code the interview runs through a tool called AskUserQuestion. In other chat tools you ask for it in words.^[1] The same interview can produce a short profile of how you write and work. Keep only lines whose absence would cause mistakes, and skip personality-type sorting, which we found no documentation for.

The second is to sketch the decision flow before building. A marketer who wants an agent to turn a campaign brief into draft emails can draw five boxes on paper first. That follows Anthropic’s explore, plan, then build order, and its page adds a limit: if you could describe the change in one sentence, skip the plan.

The third is a skill that builds skills. A skill is a folder with a SKILL.md file; Anthropic’s docs say only its short description sits in context, and the body loads when the skill is used.^[4] Anthropic already ships a skill-creator plugin that writes test cases and compares versions, so you may not need your own. Whichever you use, run the new skill in a fresh session, because the session that built it already knows what you meant. Make a job a skill when it repeats and needs more than a few lines; if it’s three lines, leave it in the instructions file.

Hacks for adding power: parallel agents, a browser and a second pair of eyes

Three habits add power, and risk with it. The first is running several agents at once.

A common assumption is that ten agents finish ten times faster, and it holds only when the pieces are independent. A marketer who gives three competitor pages to three subagents gets three summaries back without cluttering the main conversation. The same marketer drafting messaging from those summaries is doing sequential work, which belongs in one session. Each subagent has its own context window and sends back a summary.^[5] Our guide to agents and subagents goes deeper.

Agent teams, where teammates message each other, are the heavier option. Anthropic’s page says they’re experimental and disabled by default, that sequential and same-file work does better in one session or with subagents, and that three to five teammates is the place to start.^[6] The costs page puts the overhead at approximately 7x the tokens of a standard session when teammates run in plan mode.^[7] Run in parallel only when pieces don’t depend on each other.

The second habit is browser automation. Playwright’s docs say its MCP server works from the page’s accessibility tree, not pixels, that one tool is RCE-equivalent and should only be enabled for trusted clients, and that page-supplied tools and results are untrusted input.^[8] The decision rule is to try an export, an API or a command-line tool first, and use the browser only when nothing else exists, with a check such as rows collected against rows on the page.

The third habit adds scrutiny, and I’d adopt it first. Ask the agent to show evidence before it says it’s done, then have a second agent review in a fresh context. Anthropic says a fresh reviewer isn’t biased toward code it just wrote, and to tell it to flag only gaps that affect correctness, because a reviewer asked to find problems will usually find some.

Last are rules that run themselves. Instruction files are advisory, and for an action that must happen every time or be blocked, Anthropic points to a hook. Outside Claude Code, build an approval step so sending, paying or deleting waits for a named person. Our guide to human-in-the-loop guidelines covers how deep that review should go.

Hacks for keeping the bill down: output, model choice and context

Cost hacks get the loudest claims, so read them slowest. The most quoted is trimming output tokens. A community skill called Caveman does this by telling the agent to answer in clipped prose. Its page self-reports a 65% cut in output tokens on a ten-prompt benchmark, and says in the same place that input and reasoning tokens don’t change, that its rules add input tokens every turn, and that on work that was already terse you can lose money.^[9] Output is the visible part of the bill, and in my reading Anthropic’s costs page puts bigger levers first: context size, model choice, thinking effort and moving specialist instructions out of the always-loaded file.

Model choice is the clearest lever. That page says Sonnet handles most coding tasks well and costs less than Opus, and suggests a haiku model for simple subagent tasks, which you set through each subagent’s model field. Names and prices change, so check the current model list and treat “cheapest capable” as a test you run, not a setting you assume.

The third lever is context hygiene. Anthropic calls the context window the most important resource to manage, since performance degrades as it fills. Clear between unrelated tasks, and after two failed corrections clear and write a better first prompt.^[1] The costs page adds two more: keep the instructions file under about 200 lines, and remember thinking tokens are billed as output, so lower effort on simple tasks saves cost.

In my view the cheapest habit is the dullest: start a fresh conversation when the job changes. Our explainer on what drives your AI costs shows why a long conversation costs more with every message.

The one-job rule, and why 3 to 5 is a rule of thumb rather than a law

A popular rule of thumb says to give each agent three to five responsibilities. We found no study behind that number. Anthropic’s engineers advise finding the simplest solution possible and adding complexity only when needed, and note that agentic systems often trade latency and cost for better task performance.^[10] So the useful version is one job per agent, split when the job gets crowded, rather than a number.

Four signs show a job is crowded:

• You can’t write one definition of done, because two parts need different ones.
• The instructions file keeps growing and the agent starts ignoring parts of it.
• One part needs data the others shouldn’t see.
• When something goes wrong, nobody can tell which part failed.

This is where the persona habit falls apart. A job card replaces the backstory. A one-line role such as “you prepare the weekly pipeline update for the operations lead” is enough, and we found no evidence that a name and a history improve output, though they do spend tokens. If an agent talks to customers, say it’s an AI rather than inventing a human biography.

Decisions about people are the clearest job not to hand over. Our guide on when not to use AI at work gives a fit test, and for a first build see our practical guide to building an AI agent.

Before you paste anyone's install prompt

Some of the most casual advice in this genre is also among the riskiest: paste this prompt and it will install skills on your computer. A stranger’s prompt can contain instructions or scripts you haven’t read, and an agent that follows them runs with your access.

Anthropic’s Agent Skills documentation says to use skills only from trusted sources, meaning ones you created yourself or got from Anthropic, because a malicious skill can direct the agent to use tools or run code in ways that don’t match its stated purpose.^[11] The Claude Code skills page adds that a skill can pre-approve tools and can grant itself broad access, so review that setting first.^[4] We found no page on pasted install prompts specifically; this skills guidance is the closest primary source.

Treat web pages and tool output as input, not instructions, and check with IT, security and legal before any step touches customer data.

Pick three: a scored first-week plan

Fourteen habits is too many to adopt at once. Pick one from each job: one that defines, one that checks and one that controls cost. A good first three for most people are the done-and-bad brief, the evidence prompt and clearing between tasks. The right three shift by role:

• Operations lead: done-and-bad brief, job card, second-agent reviewer, because the work repeats weekly.
• Marketer: interview prompt, evidence prompt, cheapest capable model, because volume is high and tone needs a person.
• HR partner: glossary, approval gate, job card listing what the agent must not touch, because the work involves people’s data.
• Finance analyst: done-and-bad brief, evidence prompt, fresh-context reviewer, because every figure needs tracing.
• Solo consultant: interview-built profile, context hygiene, cheapest capable model, because no one else will catch a miss.

To choose the first job, use the matrix habit. Score each candidate 0 to 2 on five questions: does it repeat, is the finish line clear, is it easy to check, is a miss low-harm, and does it have an owner. The scores are opinions, so the matrix makes them visible and arguable, not settled.

Score the pilot on the lens we use for any AI rollout: use, persistence, then impact. I’d weight impact most heavily.

Teams can start this internally. Expert-led training shortens the learning curve and spreads the setup across a team instead of one keen person, and our corporate workshops cover that. For what to do after week one, see agent pilots worth running this quarter.

Sana Mian, Co-Founder of Future Factors AI

Sana is an AI educator and learning designer specialising in making complex ideas stick for non-technical professionals. She has trained 2,000+ learners across corporate teams, bootcamps, and keynote stages. Future Factors offers AI Bootcamps, Corporate Workshops, and Speaking & Consulting for businesses ready to adopt AI without the overwhelm.

More about Sana →