In partnership with
How do you give AI a complex task?
WTF is graph engineering?: A graph is a map of the work. A box handles one bounded job. An arrow means another box needs its output. A branch lets independent jobs run together. A gate decides whether to pass, retry, ask a person or stop.
Task -> Planner -> Worker -> several reviewers -> Synthesis -> Pass?
Then he gave the picture to Codex and asked it to write a Code Mode script that implemented the workflow and ran it.
His post ends with:
There's no step 3.
That is fair for a quick experiment. A workflow you plan to reuse still needs one more pass.
The drawing has to say what each box receives, which arrows carry data, what can run at the same time, what counts as a pass and when the retries stop.
That is the useful part of graph engineering.
I am Alex, welcome to ShortCu8 by Innov8.
Lets Dive Deep 🐰
⭐Today's Shortcut
Draw the workflow backward from Done.
Write exactly what "done" means beside it. Then work backward and add only the jobs required to reach it.
For the landing page:
Done means:
- the page opens
- it works on mobile
- the CTA button works
- there are no console errors
Now the Test box knows what to check. The Build box knows what it must produce. The Fix box knows when it should run.
Step 1: Pick one result
Choose a task you already give AI.
Examples include building a page, researching a topic, cleaning a spreadsheet or turning a transcript into an article.
Write the final output on the right side of the paper:
Done: a five-source research report with working links
Avoid drawing the agents yet. Start with the result.
Step 2: Draw the jobs backward
Ask what must happen immediately before Done.
For the research report, the answer may be Final check.
What must happen before that? Write report.
Continue until you reach the starting input:
Question -> Find sources -> Check sources -> Write -> Final check -> Done
Keep the first drawing simple. Four to six boxes are enough.
Step 3: Add the failure path
A straight line only describes the happy path.
Add one decision where the work can fail:
Check sources
|
+-- valid --> Write report
|
+-- weak ---> Find better sources
|
+----> Check again
Write a limit beside the backward arrow:
Maximum: 3 attempts
The limit prevents the agent from repeating the same failed step forever.
Step 4: Give the drawing to Codex
Upload the photo or screenshot and use this prompt:
Turn this drawing into a runnable workflow for my task.
First explain every box and arrow in simple language.
Do not start until the workflow matches the drawing.
Then implement and run it.
If a check fails, follow the return arrow.
Stop after 3 failed attempts and tell me what is blocked.
Ask me before deploying, deleting, sending or spending money.
Task:
[PASTE YOUR TASK]
Read Codex's explanation before it runs. If the explanation is wrong, correct the drawing or rename the unclear box.
That is easier than repairing a long workflow script after it starts.
Step 5: Improve the drawing after one run
Run the workflow once and watch where it becomes confused.
If it skips fact-checking, add a Check sources box.
If it says "done" too early, make the Done conditions stricter.
If two jobs use the same input and do not need each other's output, draw them as two branches. Codex can run them together and merge the results later.
Add a human approval box before actions such as deployment, sending an email, deleting files or making a payment.
Your second drawing can be smarter. Your first drawing only needs to work.
Now go and ride the graph engg. hype train
The ShortList
🛠️Cool Tools of the Week:
MoonEP: Moonshot open-sourced its high-performance communication library for distributed MoE workloads
Cohere: North Automations lets users use plain language to design automated workflows
Kimi K3: Open weights are now available
Grok 4.5: The latest model from xAI, built for coding, agentic tasks, and knowledge work.
📩 Innathe Shortcu8 engane undarunnu 👇️?
We read every reply - just reply to this email and let us know how we can improve !
Appo adutha Shortcu8il kanaam bie…👋
If you read till here, you might find this interesting
#AD1
Try the AI that knows your customers. No commitment.
Most platform evaluations start with a demo request and end three weeks later in a conference room. This one takes 15 minutes and puts you directly inside Gladly's interface — navigating it on your own terms.
See how AI surfaces real-time customer context before a conversation starts. Watch how a single conversation thread pulls in purchase history, channel history, and account details without a handoff.
No installation. No commitment. Start the interactive demo and see the platform for yourself.
#AD2
What is an EOR—and why are companies using it?
Opening entities in every country can be slow, expensive, and hard to scale.
That's why more companies are using EOR to hire globally faster.
See how Oyster helps teams hire, pay, and support talent in 180+ countries while staying compliant along the way.