Simplify everything — You don’t need a Graph

(5 minute “autonomous content creation to post” workflow below)

I know the flurry of posts on graph engineering is for the algo and influencers wanting to… influence? chasing buzz wordslike strung out context junkies— but be sensible instead of perpetuating the idea that carries on the false pretense that we all needs a graph to run our workflows.

I’m going to simplify this before the content factories make it noise.

Bear in mind credits

1. Scheduled runs are the moat

You don’t need a graph

Cleverly scheduled prompts and gated handoffs achieve the same result with vastly less token cost.

• Decompose your task,
• Gauge the effort required
• Pair the LLM to relevant areas in your task they excel at.

e.g → 2x pass Sol (easiest to decomp), → 3x pass Fable (complexity phase)→ Parallel pass to Grok / Gemini. (require media outputs)

2. Make Context Structural

You don’t need your fourth memory layer.

Structured directories are more effective than another RAG framework being ignored. Keep context structural to the projectand loop the upkeep.

— STATE.md is now canon.

This hold’s higher salience than agent memory.

3. Sequential modular skills as workflow

Triggered automations are brittle at best.

Transform a workflow you already have.

• Take a business process you’ve been doing for years.
• break it down into smaller tasks,
• turn them into skills
• handoff said task to the agents.

Pre-compiled business processes into modular -p triggered steps, with commits and PRs as the HITL Gate.

The less *(or more) confusing explanation = Imagine each skill is a stepping stone and your agent needs to get across the floor of lava….I’m gonna stop here.

4. Post-write hooks

Stricter Schemas & Agent reviews

Post-write lint's. with new agentic linters — shadcn lint being a good example — you can constrain work with custom prompts, decisions and rules directly in the linter. This one has teeth on it

I won’t nag about the other two italics.

5. Trade complexity -

*Complexity — Fabrication relationship*

5x `planned-with-docs + 2x scheduled agents’ repos is greater than parallel 5x init projects with agents.

Daily “pick one task from my codebase”, has done absolute wonder’s for my ADHD. (ADHD Harness post enroute)

Simplify…

It’s okay to do the un-glam thing.

Result-driven work > buzzwords.

Also, Karpathy is at the forefront of this industry, he’s literally paid to push, pull, plan, test and procure results through dense workflows to find better ways of working with models.

Why are you running the same framework?

Inherently, most jobs are pushing paper for the one above you. Whether you’re C-suite or mail boy, that is your underlying purpose. Be honest with yourself before copying frontier infrastructure you do not need.

Plus less traffic would be swell. Google isn’t kind to WA

Don’t be the Double-standards Guy

The one who doesn’t blame guns, but blames AI — when they’re both tools. Don’t

Don’t be afraid of AI because it’s going to take your job.

It’s going to take all our jobs, or at least huge parts of them.

Remember All those times you thought.

“I wish I had a clone of myself to do all the shit stuff”

Newsflash: it’s here.

And if you can’t recognize that, you’re making it longer for the rest of us.

5-min Social Content workflow

Schedule the following on their apps:

(Top Agent must be Gemini, → interchangeable claude/codex/grok)

• If you have Gemini spark — schedule a scan of your Notebooks to find material for content.
• Ask it to output optimally for the N platforms you choose and place the snippet in a drive folder /content/outbox/[social-platform]/
• Create a /done folder and a log.md instructing agents to log when snippets land and log when snippets are retrieved → move the completed ones into the folder

Create skill if successful(I did X, reddit, linkedin and meta) — this next step is optional

• Go to Chat-GPT — Schedule it to go into drive folder, log out 2 snippets from separate social folders
• Place the snippets in @canva on [name-of-template]

(Stop here if you don’t know what MCP is. Do 2x snippets with another LLM= Schedule your posts in Canva)

Create skill if successful(I did 2x chatgpt, 2x grok — content created and templated)

• Go on claude— get a buffer.com or composio.com api key and follow the instructions for adding mcp to it. (This works on grok/chat/claude)
• Schedule to grab the templates from canva and schedule posts on buffer for you to approve.

Create skill if successful.You can use whichever LLM the important part is splitting the load via scheduled tasks

Viola! simple workflow that works. You can totally skip the canva step too

Other useful practices

• Phased Interactive workflows
• Model-tailored workflows,
• self-refinement (human & model) loops .
• everything-about-me skill

The quality is in the content, not the quantity of impressions and don’t forget to create the repeatable skill on success!

ok now i do feel like I’m nagging.

Peace.

.ktg |Time to post useful shit i guess