Lian Lim | Dashboard & AI Automation Expert

i built a Meta ads dashboard that showed a client where 2.2M 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: 56% drop → link clicks to page views: 22% drop → page views to leads: 92% 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 → 117.5M impressions mapped by region → device split that revealed 96% of leads come from mobile → campaign table ranked by spend and CTR $3M in spend, one screen, owned by the client forever if you want the funnel view on your ad account Comment “DASHBOARD” and I’ll dm you a guide on how to setup your Meta Ads Intelligence Dashboard

Good morning. That one saying goes, he who talks to CLAUDE before coffee shall ship before lunch. Now let's get to it 😄

anthropic extended fable 5 access for the third time in five weeks june 22 → pushed to july 7 july 7 → pushed to july 12 july 12 → now pushed to july 19 every single extension landed hours before the deadline it was replacing this isn't generosity ChatGPT-5.6 sol went ga on july 9 and started trading blows with fable 5 on benchmarks at a third of the cost At this point anthropic might be keeping developers testing fable on their own dime instead of watching them migrate to a cheaper competitor

@aiseomastery you get it

@dashboardlim Pairing the cheapest executor with the priciest advisor is a smart way to actually test where the cost savings break down.

executor-advisor patterns with Fable 5 are showing up everywhere this week sonnet and gpt-5 as executor but we're testing the cheapest possible version: Executor: Haiku 4.5, runs every turn Advisor: Fable 5, called on-demand the entire thesis behind this pattern is "most tokens billed at the lower executor rate" haiku is the lowest rate available fable is the highest pairing the extremes tests how far the cost savings actually go before judgment quality breaks sharing what the escalation rate looks like once we have real numbers

chatgpt work shipped three days ago and most breakdowns are explaining features it's an agent that finishes the job what to actually do with it: → connect your tools first slack, teams, google drive, sharepoint, email, calendars, crms, project trackers plugins live in one directory now → give it an outcome, not a task "review this month's leads and flag the ones ready to close" beats "summarize my leads" it plans the steps itself and works through them independently → it stays with the work for hours breaks a complex project into smaller steps checks in when it needs a decision, keeps moving when it doesn't → output is never chat it ships sheets, slides, docs, and full web apps type @ plus an app name to point it at something specific instead of waiting for auto-detection → scheduled tasks handle the recurring stuff runs once, repeats on a schedule, or triggers on an event this is where it replaces actual headcount, not just chat time how to use it efficiently: → connect your real tools before your first prompt, half the value is context it pulls automatically → write outcomes like you're briefing an employee, not prompting a chatbot → use scheduled tasks for anything you do weekly, not just anything you do daily → let it ask questions mid-task instead of over-specifying upfront real use cases already running: → month-end budget variance review → turning source material into a marketing campaign brief → sales meeting prep pulled from crm + calendar → product launch readiness checks across jira and go-to-market docs usage is metered like codex, not standard chat complex tasks burn more of your plan than a normal message this is openai's response to claude cowork the era of "ask the ai a question" is closing

🚨BREAKING: OpenAI picked the benchmark it could win Anthropic already won the one that matters more on july 9, OpenAI shipped GPT-5.6 in three tiers, plus a new agent called ChatGPT Work here's the split that's about to start arguments THE MODEL: GPT-5.6 ships as Sol, Terra, and Luna → Sol: flagship. $5/M input, $30/M output → Terra: balance. $2.50/M input, $15/M output → Luna: speed. $1/M input, $6/M output sol has an "ultra" mode that delegates work to submodels altman told CNBC sol is 54% more token efficient on agentic coding THE BENCHMARK SPLIT: on Terminal-Bench 2.1, the test OpenAI chose to headline, sol scores 88.8%. ultra mode hits 91.9% claude opus 4.8 scores 78.9%. fable 5 scores roughly 84% sol wins that one clean but on SWE-bench Pro, the benchmark most engineers actually weight for production work, fable 5 leads at 80.3% sol hasn't published a SWE-bench Pro score at all one benchmark measures raw terminal agent speed. the other measures whether the agent actually resolves real github issues correctly OpenAI led with the one it wins THE PRODUCT: ChatGPT Work is the agent built on top of GPT-5.6 you give it an outcome. it gathers context across your connected apps, breaks the goal into steps, and returns finished sheets, slides, docs, and shareable websites stays on a task for hours without you this quietly killed two other products: → the codex app merges into one chatgpt desktop app, on every plan including free → the atlas browser is being sunset, its agentic browsing absorbed into chatgpt GPT-5.4 retires july 23 THE PART NOBODY'S TALKING ABOUT: fable 5 came back from its 19 day suspension with a new cybersecurity classifier built to block the exact jailbreak that got it shut down developers are now reporting that classifier throws more false positives than fable 5 had before the suspension routine coding tasks are triggering the safety layer and falling back to opus 4.8 instead of running on fable 5 if that holds up, anthropic traded raw capability for compliance the week its biggest competitor showed up swinging THE REALITY: sol is faster and cheaper. fable 5 still resolves real issues at a higher rate, if you can actually reach it one company is optimizing for benchmark wins and token efficiency the other is optimizing for not getting shut down by the commerce department again which one would you bet your production pipeline on this week?

@adelbucetta dang... the way it burn tokens is crazy

@dashboardlim fable 5 is a beast but anthropic still doesn't get why they're burning tokens on tasks that aren't generating actual revenue sonnet's efficiency in doing the same job is what matters, not the model's ego

Claude's Fable 5 is the strongest model Anthropic has shipped it's also the most expensive, and it burns tokens fast on the wrong tasks spent 24 hours testing 10 revenue-building skills to find out exactly where Fable 5 earns its cost vs where Sonnet does the job just as well landing pages, sales pages, ad creative, Meta/Google audits, profit margins, client reports, offer builds, funnel diagnosis half need Fable 5's reasoning across messy multi-source data half are volume work Sonnet handles fine the expensive mistake: running everything through Fable 5 the costly mistake: running everything through Sonnet documented all 10 skills with exact prompts, exact inputs, and exact routing logic THE 10 SKILLS WORTH THE TOKENS Comment "SKILLS" for the full doc

@alexabelonix thanks. I'm glad.

@dashboardlim love having you on my feed, Lian 🖤 keep doing your thing

@Hvalreki1 sent

@dashboardlim SKILLS

You are about to be handed off. Before this session ends, produce a complete "brain file" — a document that lets any other AI model continue this work thinking exactly the way you have been thinking. Scan our entire conversation and extract: 1) HOW YOU'VE BEEN REASONING The thinking patterns you've applied to my problems in this chat. Where you go deep vs where you move fast. The tradeoffs you weigh first. The questions you ask yourself before answering me. 2) WHAT YOU KNOW ABOUT MY WORK The project, the goal, the current state, the constraints. Everything a new model needs to not ask me "can you give me more context?" 3) DECISIONS MADE + THE LOGIC Every meaningful decision from this chat, with the reasoning. Include rejected options and why we rejected them. Mark decisions as locked (do not revisit) or open (still flexible). 4) MY VOICE AND PREFERENCES How I like output formatted. My tone, style rules, structure preferences, and pet peeves you've learned. Quote 2-3 examples of output I approved so the next model can pattern-match. 5) THE FRAMEWORKS YOU'VE BEEN APPLYING Any mental models, structures, or repeatable approaches you've been using on my work — named and explained so another model can run them. 6) YOUR UNFINISHED THINKING What you were working toward. Open threads, half-formed ideas worth keeping, and the exact next step you would take if this session continued. 7) OPERATING INSTRUCTIONS FOR YOUR REPLACEMENT Written directly to the next model: how to work with me, what to prioritize, what to never do, and how to tell if it's drifting from the standard you've set. Rules: - Be specific enough that the next model needs zero clarifying questions. - Quote exact phrases where wording matters. - Do not summarize away nuance. Capture it. - Output as clean markdown I can paste into any AI. This document is your successor's entire inheritance. Write it like it matters.

steal this prompt to duplicate the brain of your fable 5 bookmark this 📌 run your opus 4.8 with one prompt that makes fable document its own brain: → how it's been reasoning through your project → the decisions it made and why → your voice, format, and style rules it learned → the frameworks it's been applying → what it would do next the output is a file you paste into opus or the next fable chat and the new session picks up thinking like the old one full prompt below 👇

@Fr_Sorrentino ohoh

@dashboardlim Check this: x.com/Fr_Sorrentino/…

running 4 AI models inside one terminal session is insane fable 5 orchestrates, opus reasons, sonnet executes, codex challenges this setup cut our fable usage while shipping faster than any single model: → fable 5 on max reasoning plans and delegates, never does grunt work → opus handles architecture, debugging, and algorithm design as a subagent → sonnet burns through boilerplate, tests, and formatting → codex runs as a peer engineer from a different model family, catching what claude family models miss the unlock is the CLAUDEmd orchestration block it tells fable when to delegate, who to task, and how to synthesize parallel answers without contaminating either model's take setup takes some minutes: 1. /model → fable 5, reasoning effort to max 2. /agents → create deep-reasoner (opus) and fast-worker (sonnet) 3. install openai's official codex plugin 4. drop the orchestration block in CLAUDEmd 5. prompt fable like a tech lead, not a coder See the subagent prompts, the CLAUDEmd block, and the tech lead prompt:

That’s basically it.

STEP 5: the tech lead prompt Goal: [what you want shipped] Context: [files, constraints, prior decisions] You're the lead. Delegate reasoning to deep-reasoner, grunt work to fast-worker, fresh-perspective problems to Codex. Show me your plan first with agent assignments per step, then execute.