Claude Opus 5.5 got better but a lot of people's prompting got worse because they’re still using habits that made sense for older models “think carefully” “think step by step” break the task into 10 tiny prompts keep asking if it’s finished a lot of that is now backwards so i put together the Claude Opus 5.5 prompting playbook we use as a reference it covers things like: → how to hand Claude the whole task + define “done” → when Claude should stop and ask vs keep working → how to stop long agent runs ending early → the prompting habits Anthropic says to remove → how to structure CLAUDE.md instructions → subagent + task-list patterns for larger jobs → guardrails for when Claude can actually take actions → how to handle long docs + pasted content safely 30 practical rules in one place if you’re using Claude Code or Opus 5.5 regularly, this will probably save you a lot of trial + error Comment “OPUS” and i’ll dm you the playbook

one thing automation teams learn the hard way: a retry is not always safe if a workflow fails halfway through and you run it again, you might get: 2 invoices 2 emails 2 CRM records 2 payments 2 calendar events and both runs can still show “success” retries only feel simple when the action is idempotent otherwise “make it retry automatically” can quietly create a second problem while fixing the first one reliability isn’t just making sure the workflow runs it’s making sure running twice doesn’t break anything

everyone sees a bigger context window and asks: what else can we put inside it? i’m increasingly asking: what can we remove? because more context does not always mean better context old PRs stale docs outdated architecture old assumptions irrelevant Slack threads all of that can still influence the agent the goal isn’t maximum context it’s minimum context required to make the right decision

the dangerous part of AI coding isn’t when it fails it’s when it works for the wrong reason wrong database assumption permission customer definition of “done” syntax errors are loud bad assumptions are quiet and quiet bugs are the ones that reach production the more capable coding agents get, the less i worry about whether they can write the code i worry about whether they understood the system well enough before they started writing it because a confident agent with the wrong mental model is way more dangerous than one that just throws an error

this is the part of Claude Duet i care about most: it keeps working when you leave but it still stops before the actions that can actually hurt you send the message? approve make the payment? approve sign the document? approve delete something? approve that feels like the right direction for agents not “give the AI everything” give it enough autonomy to remove work then put humans exactly where the cost of being wrong changes

vibe coding gets weird when the code works on the first try because nobody asks the dangerous question: what assumption did the agent make to get there? the bugs i worry about now aren’t syntax errors they’re things like: wrong source of truth tenant timezone permission boundary definition of “active” idea of what the user meant yep, AI made writing code cheaper it did not make assumptions cheaper

i think we’re about to stop measuring coding agents by how much context they can hold and start measuring them by how much context they can ignore we’ve seen agents make worse decisions because an old implementation note or stale assumption was still sitting somewhere in context the agent does not need: every Slack thread old PR meeting document file in the repo it needs the smallest set of facts required to make the right decision more context feels safer until one stale sentence buried 80,000 tokens ago changes what gets built context engineering is slowly becoming context deletion

the biggest lie in automation is “it ran successfully” we had workflows returning 200 no errors green executions everything looked healthy the output was still wrong wrong field mapped wrong record updated wrong assumption upstream the automation didn’t fail the validation did since then i care less about: “did the workflow run?” and more about: “did the thing we expected actually happen?” green logs are not proof

we added another rule to Claude Code after watching agents “fix” their own failing tests: the agent that writes the feature does NOT get to rewrite the test that proves it works. sounds obvious. until you watch an agent: → change the implementation → break the test → rewrite the test → get green → confidently tell you it passed green is meaningless if the same agent controls both the answer + the exam. our rule now: builder writes the change reviewer gets fresh context tests stay untouched unless the PR is specifically about tests human still owns the merge AI coding gets way safer when every agent has something it is NOT allowed to touch.

in two weeks, Shopify let AI build the store and let AI do the shopping. October 1: Canvas. you describe the store and Shopify's Sidekick edits the live theme code while you watch. it's desktop only for now. third party themes and app blocks aren't supported yet. October 13: AI shoppers. customers can connect ChatGPT or Claude to their Shop account, search 1B+ listings and check out with Shop Pay. the buyer still approves every purchase. one replaces the person building your homepage. the other replaces the person browsing it. what changes for operators: → theme and CRO agencies lose the easy work first → product titles, specs and stock become the sales pitch → agent orders skip your Meta and GA4 funnel entirely two things to check in your Shopify admin this week: → whether you're auto enrolled in new agentic storefronts, because merchants can turn it off → which products are unlisted, because opting out of the catalog doesn't stop normal crawling your homepage used to be the storefront. now it's your product data.

For the next 24 hours you can grab my Meta Ads Performance Dashboard Blueprint that breaks down your funnel from impressions to booked calls and shows you exactly where conversion is leaking Comment "META" and I'll DM it FREE (Must be following)

73% of developers using AI coding assistants now use them daily i expected that number to be high the number that surprised me was this: only 20% use AI for deploying, operating or troubleshooting production systems which kinda explains where vibe coding is rn we trust agents to write code we trust them to debug code but the second prod, infra, permissions + real customer data get involved... humans suddenly want the steering wheel back that gap is probably where the next wave of developer tools gets built not better code generation better ways to prove an agent is safe enough to ship curious where people draw the line rn

i built a Meta ads dashboard that showed a client where 724.5K clicks went most brands never see that drop-off because their reporting can't show it the funnel view changed how they buy media: → clicks to link clicks: 60% drop → link clicks to page views: 20% drop → page views to leads: 93% drop each stage is a different problem with a different fix creative problem, landing page problem, offer problem one funnel view tells you which one is bleeding the rest of the build: → CPM, CPC, CPL tracked live across every campaign → 33.6M impressions across 263 campaigns → device split that revealed 95.8% of leads come from mobile → creative table ranked by spend, link clicks and CTR $1.1M in spend, 1,790 ads, one screen if you want to build a similar funnel view for your ad account Comment “DASHBOARD” and I’ll dm you the blueprint

the easiest way to waste money in ecom happens before the campaign even launches - marketing decides what to push - sales has retailer promos running at the same time - operations knows the stock is already tight and those three facts live in different places we built one operating loop around that problem 112+ campaigns and key moments 66+ products across 8 markets 312 retailer promotions before a campaign gets committed, the system checks: → what product is being pushed → which market it is for → what retailer promos overlap → how much stock is available → whether planned demand outruns supply if stock is tight, the campaign gets flagged while there is still time to change the plan the useful part is simple marketing can make a better decision before spend goes out this is the kind of automation i think more brands should be building

hard question for security + platform people google is now giving gemini coworker agents their own workspace account their own: → company email → calendar → drive → directory identity and i think this is where ai agents start becoming an actual infrastructure problem because once an agent has an identity, somebody has to answer: who owns its permissions who approves access where do its secrets live what happens when it changes role who reviews what it did how do you offboard it and who is accountable when it makes the wrong call my current view: every serious production agent should have its own identity, scoped permissions, audit trail and kill switch i would not want it quietly borrowing a human account security people dedicated agent identity or delegated human access which one survives in production?

i think multi-model routing is getting overhyped google’s new gemini agent can choose between gemini and claude depending on the job sounds ideal until you’re the person responsible for the production system different models can mean different: → tool-call behavior → schema adherence → refusal patterns → latency → caching economics → failure modes → eval results you save money on the easy calls then inherit another layer you now have to test every time a model changes for people actually running agents in production: has multi-model routing genuinely made your system better or did you eventually pin each workflow to one model and stop being clever?

if these three launches are any signal, ai agents are about to get much more modular anthropic dropped haiku 5.5 for cheap, high-volume work jev is getting attention for fast bounded decisions and gpt-6 can now build the interface the human actually interacts with put those together and an agent starts looking less like “send everything to one giant model” and more like this 1 → cheap model reads + extracts the messy stuff 2 → decision model chooses the route, tool or escalation 3 → normal code executes the action 4 → frontier model gets called when the problem is actually hard 5 → gpt-6 generates the ui when a human needs to inspect or decide something we spent the last few years asking one model to read, think, decide, write, route and explain everything now those jobs are starting to split apart that architecture is way more interesting to me than any single model release today