Advanced AI-SDLC for Students.
Advanced tips and howtos for SDLC from day-to-day working with SDLC. Not for students only — for anyone who ships code with an agent.
The name is a little misleading. Let's fix it before we start.
Should be aware of the underlying concepts — not just how to ship.
Can't afford to start with vibe-coding. Need a minimal theoretical understanding first — then the vibe-coding compounds.
Should be aware of SDLC too.
The shape of the pipeline — plan → write → test → deploy → observe — hasn't changed. What's changed is who runs each step.
A tweet from Boaz Ha'Nasich got a lot of rage. Good.
Disclaimer: Boaz speaks only for himself.
Having said that — the tweet got a lot of «rage».
Having said that #2 — it's good to hear different opinions before we settle on the «right» way.
Basic SDLC is already solved. We're at the 80% of it.
Plan, write, review, test, deploy — the baseline motions have all been automated at least once. The interesting 20% left is where this talk lives:
· the seams between tools
· the operator's judgment calls
· the boring engineering that makes agents actually usable
The vocabulary you'll trip over. Pick the ones that ship.
Also relevant: AI fatigue — the reason half these words will be dead in a year.
Turn every step of a research loop into an MCP tool.
Lite NLP research over Twitter twits — hot topics, trends, samples. Steps:
What if we wrapped each step as an MCP tool? Then Claude does the whole research loop — picks the query, runs the topic model, reads results, iterates.
Claude becomes the engine that does the operations, not just the assistant that suggests them.
Concretely: a small set of domain tools that Claude can compose.
Data scientist writes the NLP functions.
Data analyst / marketer uses the MCP to compose them.
OR — Claude becomes the engine that runs the operations. That's the interesting version.
(also — kinda «old news» now)
A partial list of things that just work now.
Pareto 80/20. Stick to one. Learn its mechanics.
A year ago you couldn't imagine what Sonnet 4.6 or GPT 4.3 would ship. Whichever you picked, you'd have been fine.
Better than moving CC → Codex (or the reverse) weekly: learn something durable — like the «async programming model» on the next slide.
The market re-forms around the tools every six months.
Mobile & web interface for Claude Code and Codex.
Question: would they be accepted to Summer 2026 — or would the incumbents have absorbed the surface by then?
ycombinator.com/companies/omnara
Paid for Cursor in mid 2023. Stopped after a few months.
Then Cursor skyrocketed. Then came Claude Code.
The saga continues. Lesson: don't bet the company on any one tool. Bet on your ability to switch.
Set and forget. The primitive that changes how you plan work.
Tools = superpowers for the agent. Not reinventing the wheel.
Tools let the agent:
Classical NLP is mostly solved. Don't make the LLM reimplement Spacy. Wrap Spacy as a tool and let the LLM call it.
Same for everything with a good library: image processing, geospatial, graphs, DB, HTTP, shell.
Wrap the tools you love. Give them to the agent.
github.com/jehna/humanify
What if we let Claude Code control this lib and de-obfuscate JS in a loop?
Update: seems like CC can do it itself. See @WeizmanGal.
What if we let Claude Code do RE on binaries?
What if we connect Ghidra to Claude Code?
github.com/LaurieWired/GhidraMCP
— already exists.
Writing good prompts is not «prompt engineering.»
Describing the LLM's role — «think like a senior architect.»
Few-shot examples — already automatable.
Trick incantations — «think step by step.»
Reproducible on demand. Trainable into the model. Not durable.
Distilling your domain into the LLM — the rules only you know.
use uv venv · use dockers · no GitHub repos under 100 stars · unit tests → e2e tests → stress tests
The operator's real job. Teaching what you know.
Don't. Let Claude Code write it. Then run it. Then debug it.
The recursion is the point. Every agent you write should have been written by an agent.
Almost every dev process can be mimicked. Knowing it exists is the hard part.
@Amit_Mandelbaum
Reproducible with perplexity + Twitter search. But you have to know Remotion exists.
Yahav Fuchs · LinkedIn
Reproducible by asking Claude Code. But you have to know about it first.
The 10× comes from discovery, not execution. Read widely. Follow builders. The multiplier is downstream of what you know is possible.
Not a problem — and also unsolved.
Practical advice: don't hand-curate. Let the pile grow. Add a «compare-history» skill when it starts hurting.
Models are trained on their harness. Follow the harness's conventions.
The training data for Claude Code and Codex heavily includes their own harness: the tools, the shell environments, the RL sandboxes.
What this means for you: use the conventions. A CLAUDE.md at the repo root works better than any prompt trick, because the model was trained expecting one.
Very simple. One prompt. Ten parallel projects.
Use chrome and get several RAG projects from github. Take 10–20 most common projects. Check licenses; don't use projects with non-permissive licenses. Clone locally. For each project, add a ralph task: · review main business logic · understand quirks this project solves · understand high-level architecture · understand some non-trivial logic Create 2–3 ralph tickets per project: · analysis · running project in container Then summarize all we know. Create combined RAG system with code from other systems. Evaluate. You have OpenRouter api key in env vars. Use ralph skill to convert to json. Run at most 2 tasks in parallel.
This works. I have tested it. Not as simple as it looks. But it absolutely 100% works.
Offer suggestions. Agentic retrieval. Read the news. Try the news.
Two follow-up moves after the first pass:
The multiplier is on the loop, not the model. Keep the loop honest.