Skip to main contentSkip to navigation

A domain track

Claude & AI for Engineering

Backend and frontend engineers, DevOps and QA — teams already using AI every day who have hit the ceiling of what one prompt at a time can do.

35 topics

Prompting one file at a time works until about the tenth. After that you are pasting the same instructions again, and someone reads every output from scratch because nothing tells you whether it is right. This workshop is for engineers who already use AI daily and want to stop typing requests and start designing loops: what the agent does, what checks it, and when it stops.

The failure mode is not a single wrong answer. It is an agent that burns a whole night on a task that stalled at attempt three, or drifts after a dozen iterations and rewrites code that was never broken. So every loop here starts with a stop condition and a run ceiling. The reviewer agent gets read-only access, which keeps it improving the work instead of gaming the test, and the morning report says what finished, what stalled and what is waiting for you.

Everything runs on your repository: the files with no tests, the CI build that keeps going red, the module nobody documented. Several agents work at once, each on its own isolated branch, and a context audit shows what every turn costs in tokens. Each participant leaves having run a loop with a real stop condition and opened a PR that came out of it.

Online course

Start online, at your own pace

The material we teach in the room, recorded slide by slide. You go through it when it suits you, from any screen.

Claude & AI for Engineering: the online course is on its way

Leave your email and we'll write once, the day it opens.

What happens to your details

Giving us these details is your choice, and no law requires it. Without the fields marked required, though, we can't finish what you came here to do.

What for: To email you once, the day the online course for your role opens. Nothing else.

Who gets it: Wider Group only. Your address sits with our infrastructure providers, Supabase and Vercel. Our email provider, Resend, sends us a note when you sign up. All three run servers outside Israel.

You can ask to see what we hold on you, correct it, or have it deleted, any time. Write to shahar@wider-group.com and we answer within 30 days. The full detail is in the Privacy Policy.

Rather do it live? See the live workshops

Today's challenges

  • Prompting works until the tenth file, and from there it is pasting the same instructions over and over
  • No objective way to tell whether the output is right, so a person re-reviews all of it anyway
  • Runs that go all night with no stop condition, burning budget on a task that stalled at attempt three
  • An agent that drifts after a dozen iterations and starts changing code that was never broken
  • Nobody can say what the agent did last night, so trust erodes until the team quietly stops using it

Example use cases

  • A loop that runs until every file has a passing test, and stops itself at a run ceiling
  • A reviewer agent with no permission to touch the code it grades, so it improves the work instead of the test
  • Agents running in parallel on isolated branches, so ten files move at once without collisions
  • A nightly run that reports each morning what finished, what stalled, and what is waiting on you
  • A context and cost audit showing exactly what you pay for on every turn, and what can go
  • An agent that diagnoses a failing CI build and submits the fix as a PR

What we teach here

Intro

A live opening session built to shift the room, not lecture it: real demos of what Claude can already do inside your workflows, run in front of the group. Participants leave with a working sense of where AI helps and where it doesn't, and enough curiosity to start using it that afternoon.

AI & LLM Basics

What tokens, context windows and hallucinations mean, explained without the math and without the hand-waving. Once a team understands why a model sometimes gets things wrong, it stops either trusting it blindly or dismissing it after one bad answer. Every other module in the program builds on this one.

Claude Intro

A new AI tool comes out every week, and you don't need to chase them. We map the main ones and what each does best, then explain why we spend the day in Claude. After that we set it up together: privacy settings, the one capability to switch on, and a tour of the screen. Face to face it is a live demo; in the online course it is a recorded tour.

Claude Skills

Claude Skills turn a procedure your team already does well into something Claude can repeat on its own: the same checklist, the same tone, the same edge cases covered every time. You build it once with the person who actually knows the process, and from then on every teammate gets that expertise on demand. For teams where quality depends on one or two people, this is how that stops being a bottleneck.

Prompt Engineering

Structure, context and role framing: the craft of asking in a way that gets a usable result on the first try instead of the fifth. Participants practice on real prompts from their own work, not toy examples, so what they take home already fits the job. Every other module assumes this fluency, which is why it comes early.

Claude Projects

Claude Projects hold your documents and instructions in one persistent workspace, so every conversation starts already knowing your company, your terminology and your rules. Teams that set this up once save the ten minutes of re-explaining context at the start of every session, every day. For anyone running the same kind of task repeatedly, this alone changes the math.

Design systems: your brand inside Claude

Without a design system, Claude decides how everything looks, and it comes out different every time. The session shows what a design system holds, how every slide, dashboard and app is built on it, how to make one from your website and a deck you already have, and how to keep several and switch between them.

Artifacts: from answer to deliverable

The difference between an answer and a deliverable. The session covers when something should become an artifact, every kind there is today, how they all come out on your design system, the four ways to change one, and what you do with it: connect it to apps, connect it to files, share it as a live place, or export a snapshot.

Agentic AI vs. Chat

The shift from asking Claude a question to handing it a multi-step task it completes on its own, checking in only where it matters. Participants learn what actually counts as an agent, when handing over work beats doing it turn by turn, and how supervision needs to change once Claude is running unattended. Get this wrong and you either babysit an agent that didn't need it or trust one that did.

Agentic workflows in the Claude chat

In the Claude chat you can hand over a whole job instead of asking a question. Claude plans the steps, searches the web, reads your files, runs code and hands back a finished document, spreadsheet or deck. On a big task it can split the work between several sub-agents running at once. Participants learn which work is worth handing over and how to brief it the way they'd brief a colleague, then practice staying in control while Claude works. They finish by turning a weekly chore into a scheduled task that runs on its own.

Routines in Claude

Tasks Claude runs on its own, on a schedule: a morning brief from your calendar and inbox, a weekly report, a competitor watch. Participants learn to tell a routine from its schedule, connect only the connectors it needs, and test every routine before it runs alone. They leave with one routine that works from the next morning.

Connectors & MCP

Connecting Claude to the systems your team actually runs on: drives, CRMs, databases, through connectors and MCP, with the permission boundaries covered before the connection ever gets made. Participants leave knowing exactly what to expose and what to keep out of reach. This is the module that turns Claude from a smart typewriter into something acting on real, current data.

Claude Code

Real engineering workflows built agent-first: how a task moves from a prompt in the terminal to reviewed, shipped code. The module is built for developers, but the data-analysis half works just as well for a non-developer who lives in spreadsheets and needs that same rigor applied to a script. Teams leave with a working setup running in their own repo.

CLAUDE.md

CLAUDE.md is a plain text file Claude Code reads at the start of every session, so whatever you'd explain to a new hire on day one, you write there once. Participants learn what goes in it and what stays out, when it's read, and where it lives: the organization, you, the project and a subfolder, all loading together. They leave with a draft from /init and one rule for growing it: what you corrected twice goes in the file.

Agents for non-engineers

How to build an agent without writing code: it gets a job description, a badge that limits what it can touch, and a bottom line it hands back. Participants see why the work goes to a separate room, take an agent file apart, and fill in the one-page charter that turns into the agent's file. They learn to hire for a role, not a task, and to tell when a job needs an agent rather than a skill. They leave knowing the hard part is the charter, and that the charter is their own expertise, written down.

Plugins: the department in a box

A plugin carries the agents, skills and hooks a team built, so another team installs the same department instead of rebuilding it. Participants see what goes in the box and what stays home: CLAUDE.md and the permission rules belong to each team. Then they ask Claude to build one, try it in a fresh session, and share it through a marketplace.

Context Management

How context windows, memory and cost interact, and what to do about it before a long working session starts drifting or the bill surprises someone. Participants learn to spot when Claude is still hauling a task that should have been closed an hour ago, and how to reset without losing what mattered. A short module with an outsized effect on both output quality and the invoice.

Loop Engineering

Designing autonomous agent loops that intake a task, execute it, review their own work and recover from failure without a person watching every step. This is where the program moves from using Claude to building systems that use Claude. Teams walk out with a loop design they can actually run.

Agent Teams & Orchestration

Orchestrating several agents at once: how to split work into specialist roles, run pieces in parallel, and integrate the results into one coherent output instead of five conflicting drafts. This module targets teams already comfortable with a single agent and ready to scale up. It closes the gap between a useful demo and a system a team can rely on.

Safe AI at Work

The rules every employee needs before their first real session: what data is safe to share, what isn't, and what org-level guardrails should exist so individual judgment isn't the only safeguard. This module is usually the one legal and IT ask for by name. It replaces vague anxiety about compliance with a clear, specific policy people can actually follow.

Human in the Loop

Where people belong in an AI-assisted workflow: which steps need a review gate, which need an approval, and how to calibrate trust so oversight isn't either everywhere or nowhere. This pilot module gives teams a concrete framework instead of a gut feeling about what to check. Get the calibration right and quality goes up without slowing anything down.

Security & Privacy

Permissions, prompt injection and data boundaries: what makes an agent that can act also one that can go wrong, and how to run it anyway without carrying the risk. This pilot module is built for organizations moving agents past read-only demos into things that write, send and change data. It's the module security teams should sit in on.

Token Management

Treating tokens as a real budget: measuring current spend, predicting where it grows, and cutting it without cutting the quality of what Claude produces. This pilot module is built for teams past the experimentation phase, where AI cost is now a line item someone has to defend. Finance and engineering leave with the same numbers, which usually hasn't happened before.

Claude Sub-agents

Splitting a job across specialized sub-agents: when spawning a new one is worth the overhead, what to hand it, and how its output comes back into the main thread cleanly. This pilot module is the natural next step after agent teams, aimed at engineering and product teams building more complex systems. Participants leave with a decision rule, not just examples.

Claude Plugins

Extending Claude with plugins: how to evaluate one before installing it, what to trust, and when the right move is building the one your team is missing instead. This pilot module keeps the scope practical rather than a tour of everything that exists. Teams leave able to tell a useful plugin from a risky one.

Agent Evals & Testing

Coming soon

A syllabus item still coming: systematic evaluation of agent output, including regression tests, so a team can tell whether an AI feature works and keeps working after the next change. Participants will build evals for their own agents, not a generic benchmark. Built for teams shipping agent-based features who currently have no way to catch a regression before a user does.

Observability & Debugging

Coming soon

Coming soon: tracing and logging agent runs so a failure is something you can debug, not a mystery you re-run and hope goes better next time. Engineering teams will learn to build the observability layer before they need it, not after a production incident forces the question. It's the module that turns a strange agent glitch into an actual root cause.

Parallel Agents & Version Control

Coming soon

Coming soon: running many agents against one codebase without them stepping on each other, covering worktrees, branch strategy and merge discipline built specifically for agent-scale parallelism. This is for engineering teams already running more than one agent and starting to feel the coordination cost. The syllabus targets the failure modes teams hit first: stale branches, silent overwrites, conflicting assumptions.

Graph Engineering

Coming soon

Coming soon: building a knowledge graph of a codebase or domain so an agent stops re-exploring the same territory every time it's asked something new. Participants will map one of their own systems into a graph an agent can actually query. Aimed at teams whose agents spend more time searching than working.

Harness Engineering

Coming soon

Coming soon: building the harness around the model itself, tools, permissions, hooks, and the scaffolding that turns a capable model into a dependable agent. This syllabus item goes deeper than any other module into the infrastructure layer, for teams building agents as a product rather than using one off the shelf. Expect to leave with a harness design, not a checklist of terms.

Machine Learning Basics

Coming soon

Coming soon: the foundations under the products everyone's using, training, fine-tuning and embeddings, taught for the decisions a non-researcher actually has to make. This is not a path to becoming a data scientist. It's enough grounding to evaluate a vendor's claims or a build-versus-buy question without taking their word for it.

Data Management

Coming soon

Coming soon: preparing and governing the data an AI system actually works from, structure, access and quality, before automation gets layered on top of it. This syllabus item is for teams that already tried an AI project and hit a data problem they weren't expecting. Skipping this step is the single most common reason automation projects stall.

AI Policy & Governance

Coming soon

Coming soon: writing the AI policy an organization actually needs, usage rules, risk tiers and rollout governance, instead of a generic template nobody follows. This syllabus item is for leadership and compliance stakeholders shaping how AI gets adopted company-wide. Built to produce a document teams can point to, not a discussion that goes nowhere.

RAG & Knowledge Bases

Coming soon

Coming soon: retrieval-augmented generation done properly, chunking strategy, embeddings, and answers traceable back to a specific source rather than a plausible-sounding guess. This syllabus item is for engineering and product teams building a knowledge feature on top of Claude, not just using it in a chat window. Participants leave with a working RAG setup on their own documents.

Claude API & SDK Basics

Coming soon

Coming soon: building on the Claude API and Agent SDK, first integration, tool use, streaming, and the cost controls that matter before something ships to production. This syllabus item is for engineers moving from using Claude manually to embedding it inside a product. Expect a working integration by the end, run against real endpoints.

What lands in the repo

Four setups that go into the next sprint, each with a clear limit on what the agent may touch.

A loop that knows when to stop

A configuration that runs until every file has a passing test, then halts itself at a run ceiling or a token budget you set in advance. The limits are yours, and no loop gets to decide on its own to keep going.

A reviewer with no write access

An agent that grades code another agent wrote and has no way to change it. Its comments come back to you, and the person who approves the merge is still an engineer.

Parallel agents, isolated branches

Ten files move at once, each on its own branch, with no collisions. You end up with one PR per branch and review them like any other PR before anything merges.

A morning report and a cost audit

A nightly run that sums up what finished, what stalled and what waits on you, next to an audit of what each turn costs. An agent that diagnoses a failing CI build submits its fix as a PR, and you decide whether it merges.

Questions people ask

Is there a Claude Code workshop for engineering teams?

Yes. Wider Group runs a workshop for engineering teams, on site or remote, in Hebrew or English. It works on your own repository and CI setup, not a demo project, and focuses on loops with stop conditions, a read-only reviewer agent and parallel agents on isolated branches. It is quoted per team, with no public price list.

How do I stop an AI agent from running all night on a stuck task?

Every loop gets a stop condition and a run ceiling or token budget before it starts. Wider Group teaches engineering teams to set these limits up front, give the reviewer agent read-only access so it cannot game the test, and read a morning report that says what finished, what stalled and what waits for a human decision.

Is there an online Claude course for developers?

Not yet as its own version. Wider Group's online course is self-paced, in Hebrew and English, and not open for purchase yet, with a waitlist. Role versions exist for HR, finance, marketing and product, and engineers start with the general one. For work on your repository, the team workshop is the right fit.

Tonight's run

Tailoring means the loops run against your repository, your CI configuration and your token budget instead of a demo project. Together we set what an agent may touch and what stays off limits, so you leave with configurations you can switch on tonight. Tell us how many engineers and which repository, and we'll come back with a quote.