February 2026·advanced ai-sdlc
01 / 20
nav·F full·T theme
WorkshopFebruary 2026

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.

00Naming issues

The name is a little misleading. Let's fix it before we start.

Students

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.

Non-students

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.

01Motivation

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.

02Why advanced?

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

03Landscape · the buzzwords

The vocabulary you'll trip over. Pick the ones that ship.

AGENTS.md / CLAUDE.md / skills.md
The agent's memory files — global rules, per-project context.
Ralph Wiggum
A pattern for running many autonomous agents in parallel worktrees.
OpenCode
Open-source coding-agent framework.
MCPs / Tools
The API layer that gives an agent superpowers beyond its context.
Multi-agent orchestration
Multiple agents debating, delegating, verifying each other.
Spec-driven development
Write the spec first; let the agent implement + test against it.
Context engineering
Feeding the agent what it needs to know — the new prompt engineering.
Windsurf · Kiro · Cursor · Codex · Claude Code
Pick one. Learn its mechanics. Don't shop between them weekly.

Also relevant: AI fatigue — the reason half these words will be dead in a year.

04MCP-ing NLP research

Turn every step of a research loop into an MCP tool.

Lite NLP research over Twitter twits — hot topics, trends, samples. Steps:

Scrape Twitter via apify
Topic model with LDA / BERTopic
Manual review in Jupyter
Iterate

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.

04bThe API surface

Concretely: a small set of domain tools that Claude can compose.

scrape_by_keyword / scrape_by_user
sentiment_analysis over a text batch
cluster a list of texts
find_nearest_neighbours of a given tweet
extract_keywords
sample_from_cluster

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)

05What is already solved

A partial list of things that just work now.

Browser control with LLMs
Not perfect. Usable.
Reverse engineering (JS, Ghidra)
De-obfuscation, binary triage — assisted, not autonomous.
Customer support chatbots
At varying levels, obviously.
Document processing
OCR + extraction + structured output.
Marketing automation
Copy, variants, targeting.
Search + RAG
Table-stakes now.
Transcription + speech
Whisper-family for most languages.
Logs analysis
Anomaly detection, root-cause narratives.
06Claude Code or Codex?

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.

How to add an MCP
How to add a skill
How to resume a previous chat (yes, really)
07Two use-cases worth watching

The market re-forms around the tools every six months.

Omnara (YC Summer 2025)

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

Cursor (Personal saga)

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.

08Async programming model

Set and forget. The primitive that changes how you plan work.

Set a task with a clear success condition
Agent runs in the background
You come back later and check the diff / logs / PR
Agents are very autonomous — more than most people expect
09Tools

Tools = superpowers for the agent. Not reinventing the wheel.

Tools let the agent:

Fix / iterate where the model alone can't
Run classical logic the model would guess badly at — Spacy, NLTK, sklearn
Browser tool — actions on your behalf
The most basic: run Python, run tests

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.

09bTools · leveling up

Wrap the tools you love. Give them to the agent.

De-obfuscate JavaScript

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.

Reverse engineering binaries

What if we let Claude Code do RE on binaries?

What if we connect Ghidra to Claude Code?

github.com/LaurieWired/GhidraMCP

— already exists.

10Writing good prompts (not «prompt engineering»)

Writing good prompts is not «prompt engineering.»

«Prompt engineer»

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.

Writing good prompts

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.

11How to write an agent

Don't. Let Claude Code write it. Then run it. Then debug it.

Talk to Claude Code. Watch it solve the task by hand.
Ask Claude Code to write the agent, using your framework of choice.
Ask Claude Code to run the agent on a few samples.
Debug the process — but tell Claude Code to debug. Not yourself.
You can debug too. Mostly you should steer: what to try next, when to stop.

The recursion is the point. Every agent you write should have been written by an agent.

12Knowing has a great value

Almost every dev process can be mimicked. Knowing it exists is the hard part.

Sample 1

@Amit_Mandelbaum

Reproducible with perplexity + Twitter search. But you have to know Remotion exists.

Sample 2

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.

13Too many markdown files

Not a problem — and also unsolved.

With agentic coding, Claude Code uses .md files as memory
Uses them as pointers — jumps between files
It knows when a markdown refers to an old version
~
Harder to find the relevant info as the pile grows
~
Pruning process? Treating as long-term memory? — unsolved

Practical advice: don't hand-curate. Let the pile grow. Add a «compare-history» skill when it starts hurting.

14The importance of best practices

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.

Reference
@daniel_mac8
15How to write a SOTA RAG system

Very simple. One prompt. Ten parallel projects.

operator prompt · build a SOTA RAG system
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.

15bHuman in the loop · close

Offer suggestions. Agentic retrieval. Read the news. Try the news.

Two follow-up moves after the first pass:

«Read this link and implement. See evals after.»
«Read latest news on Twitter/LinkedIn about RAG and memory. Try to implement the ideas.»

The multiplier is on the loop, not the model. Keep the loop honest.