Learn Claude Code, from searching a codebase to building agents

APIClaudeSam SwiggyInshortsDelhiveryRazorpayPhysicsWallahNykaa

This free course has 18 chapters, and each one works through a problem at an Indian company in drawings that change as you scroll. In chapter 1, Claude Code finds the one delivery rule that has to change in Delhivery's codebase of about 1,800 files.

Start chapter 1: Built-in tools

The chapters start with what you use on your first day, such as searching code and giving Claude Code your team's rules. They end with agents that work through long jobs on their own.

This page follows Sam, a developer who has never built anything with Claude, through the course one skill at a time. In the drawings, Claude is the octopus in the tank.

Sam finds one rule in code he has never seen

APIClaudeSam routing_engineabout 1,800 files api/legacy/v1/allocation/tests/ context window 95% full6.3% full GlobGrepReadEditWrite cod_rules.pystill unreadafter 36 files assign_centre(found in 3 files3 of about 1,800files read Editthe text appears4 timesWriteline 15 changedClaude CodeGrep3 paths

Delhivery wants to change who handles its most expensive cash-on-delivery parcels. The rule that decides this is somewhere in a ten-year-old codebase of about 1,800 files, and Sam asks Claude Code to find it.

If Claude reads the code folder by folder, everything it reads stays in its working memory for the session, which is called the context window. After 36 files that memory is almost full, and Claude still has not reached the file with the rule.

So Claude searches first, the way you would search a shared drive. The search points to 3 files, Claude reads only those, and most of its memory is still free.

Claude then changes the rule. Its first attempt is refused because the same line appears four times in the file, so it rewrites the whole file, and the change Sam reviews is one line long.

Sam teaches Claude Code how Razorpay's payments team works

APIClaudeSam CLAUDE.md40 lines,read at the start ofevery session a testing note inthirty folders one testing ruleloads only for thefiles it names hookblocks an edit thatlogs a full cardnumber PAY-3101PAY-3102PAY-3103PAY-3104 edit straight away plan mode48 files in threeservices, and twodesigns for receipts > /payments-review payments-review.mda saved prompt SKILL.mda command as a folder from one laptop to all 40 Claude CodeClaude CodeClaude Code

Razorpay's engineers were typing the same team rules into every chat. Sam writes them once in a short file called CLAUDE.md, which Claude Code reads at the start of every session, and adds a check that stops Claude from writing card numbers into the logs.

Small jobs go straight to Claude. For a change that touches 48 files, Claude first writes a plan for Sam to approve, because the last time it skipped that step, 20 files had to be fixed again.

Sam turns the team's most used instructions into commands that anyone can run by typing a slash and a name, and shares them with all 40 engineers.

Sam connects Claude Code to the team's tools and has it check every change

APIClaudeSam jiragithubledger .mcp.jsonteam servers go here tokenreached 40 laptopsand the git history PAY-2817Claude reads the ticket itself PR #4127a skipped question scores -1 instead of 0 writing sessionapproved it new sessionbug on line 88 4 tests in place of 142 of them fail on line 88 Claude Codereview job

Razorpay's engineers were copying tickets and records into Claude by hand. Sam connects Claude Code directly to the team's ticket tracker, code and payment records, and keeps the passwords out of the settings file that everyone shares.

PhysicsWallah has Claude review every change to the code that scores its mock tests. A bug that gave a skipped question minus one mark got past the session that wrote it, and a fresh review session caught it.

Sam builds a Swiggy help agent and sets limits it cannot skip

APIClaudeSam Support Where's my order? It's been 70 minutes Your rider's GPShasn't moved for 25minutes. I'verefunded ₹598. orders#5521 Thali 349 + Roll 249₹598 ON THE WAY rider idle 25 min payments order 5521the customer's order order 5512another customer's ₹640 refunded verification skipped:12 of 100 chats PreToolUsehookruns before the tool runs under ₹1,000: the agentrefunds it ₹1,000 or more: a person form₹4,800 customer careSwiggy's code agent loop chatget_order, track_riderresultsissue_refundresultSwiggy's codeissue_refund ₹4,800

A Swiggy customer writes, “Where's my order? It's been 70 minutes.” Sam writes the program that sits between the customer and Claude.

Claude knows only what the program sends it. It cannot see Swiggy's orders or where the rider is.

So Claude asks the program to look up the order and the rider. The program does both and sends back the answers: the order cost ₹598, and the rider has not moved for 25 minutes.

Claude then asks for a refund, and the program issues it. Claude asks, the program acts and Claude reads the result, again and again until Claude has nothing more to ask, and this repeating exchange is called the agent loop.

Claude's reply goes to the customer: “Your rider hasn't moved for 25 minutes. I've refunded ₹598.” Chapter 7 builds this agent step by step.

Swiggy then opens the agent to customers who are not logged in. Claude was told to check who it is talking to first, but it skipped that check in 12 of every 100 chats, and one customer got a refund on someone else's order.

An instruction to Claude can be skipped, so Sam puts the check in Swiggy's own program, where it runs before every refund. A check of this kind is called a hook.

Small refunds go through on their own. When Claude asks for a ₹4,800 refund, the hook stops it and passes the case to a person in customer care.

Sam gets clear answers from Claude and checks them

APIClaudeSam bot comments in one month 412 of 610 resolved without a change .github/review-prompt.md report: bugs and security problems skip: style and what the linter checks a severity scale with code examples three worked examples invoiceRangrez kajal12 at ₹655 extract_price_list "rate_inr":555655 "quantity": 12"stated_total": 17700 retry with the exact errorline items add up to 16,500, not stated_total 17,700 1 2 3 4 5 Message Batches APIhalf the price, results can take up to 24 hours4,000 requests, each with a custom_id review jobNykaa's codeNykaa's codebatch4,000 requests

PhysicsWallah's automatic code reviews left so many unhelpful comments that engineers ignored two of every three. Sam rewrites the review instructions to say exactly what to report and what to skip, with worked examples.

Nykaa turns suppliers' PDF price lists into product pages. Claude fills in a fixed form for each product and Nykaa's program checks the numbers. When Claude misreads a price, one retry that points to the mistake fixes it.

Before its Pink Friday sale, Nykaa receives thousands of supplier PDFs a day. Work that nobody is waiting for goes through a slower batch service at half the price.

Sam builds a research team of Claude agents for Inshorts

APIClaudeSam searchbrief document-analysisbrief synthesisbrief reportbrief 18 tools, one line each analyze_contentAnalyzes the given content and returns key information. analyze_documentAnalyzes a document and returns key information. the November bulletin,a PDF 19 of 20 to the right tool search_newshospital visits, November 2026 30 s result A[ ]reads asfound nothing result BisErrorThe news archivedid not answerwithin 30 seconds. Inshorts' codecoordinatorbriefInshorts' codebulletinInshorts' codeisError

An Inshorts editor asks whether Delhi's air this November is better than in earlier winters. The sources are too long for one Claude to read at once, so one Claude splits the question and gives each part to a helper Claude with its own short brief.

At first Claude kept sending a government bulletin to the wrong tool, because the descriptions of two tools read almost the same. After the descriptions are rewritten, 19 of 20 requests go to the right tool.

When the news archive did not answer, the search tool sent back an empty list, and the report then said there was no news. Now the tool says plainly that the archive did not answer.

Sam keeps long and large jobs accurate

APIClaudeSam one call per file 1 2 3 4 5 6 7 8 9 10 11 12 13 14 one integration callhow the files connect resume--resume edd-test-plan fork fresh startsummary 70 turns case facts order 7712₹860promised 7:10 pmdelivered 7:52 pm sent fresh with eachcustomer message asks fora person handed overstraight away,before any checks a person case factsorder 7712, ₹860 document-analysisreads the bulletins AQI 352government monitorsAQI 418private sensor networkhospital visits86% correct Delhivery's codeClaude CodeSwiggy's codeSwiggy's codeInshorts' codecoordinator

One review of a 14-file change at Delhivery missed a time-zone bug. Reviewing each file on its own and then checking how the files fit together found it.

A Delhivery engineer works with Claude Code over several days. Naming each session lets them reopen the right one, branch it into two, or start fresh from a summary.

A Swiggy customer chats for over an hour and a half about three orders, and the agent starts mixing up amounts and times. Swiggy's program keeps the key facts separately and gives them to Claude with every new message.

The agent also hands chats to people. A customer who asks for a person is handed over straight away.

Inshorts wants its agents to read air-quality reports. Every number is saved with where it came from, so when the government's monitors and a private sensor network disagree, the story can report both.

The five areas of the Claude Certified Architect: Foundations exam

APIClaudeSam Agentic Architecture & Orchestration27%chapters 7, 8, 12, 15 and 16Tool Design & MCP Integration18%chapters 1, 5, 13 and 14Claude Code Configuration & Workflows20%chapters 2, 3, 4 and 6Prompt Engineering & Structured Output20%chapters 9, 10 and 11Context Management & Reliability15%chapters 17 and 18

Each chapter covers one of the five areas of the exam. Agentic Architecture & Orchestration is 27% of the exam and is covered in chapters 7, 8, 12, 15 and 16. Tool Design & MCP Integration is 18%, in chapters 1, 5, 13 and 14. Claude Code Configuration & Workflows is 20%, in chapters 2, 3, 4 and 6. Prompt Engineering & Structured Output is 20%, in chapters 9, 10 and 11. Context Management & Reliability is 15%, in chapters 17 and 18.

New chapters open each Wednesday until Wednesday 16 December, and the list below gives each chapter's date.

Start chapter 1: Built-in tools
APIClaudeSam SwiggyInshortsDelhiveryRazorpayPhysicsWallahNykaa routing_engineabout 1,800 files api/legacy/v1/allocation/tests/ context window 95% full6.3% full GlobGrepReadEditWrite cod_rules.pystill unreadafter 36 files assign_centre(found in 3 files3 of about 1,800files read Editthe text appears4 timesWriteline 15 changedClaude CodeGrep3 paths CLAUDE.md40 lines,read at the start ofevery session a testing note inthirty folders one testing ruleloads only for thefiles it names hookblocks an edit thatlogs a full cardnumber PAY-3101PAY-3102PAY-3103PAY-3104 edit straight away plan mode48 files in threeservices, and twodesigns for receipts > /payments-review payments-review.mda saved prompt SKILL.mda command as a folder from one laptop to all 40 Claude CodeClaude CodeClaude Code jiragithubledger .mcp.jsonteam servers go here tokenreached 40 laptopsand the git history PAY-2817Claude reads the ticket itself PR #4127a skipped question scores -1 instead of 0 writing sessionapproved it new sessionbug on line 88 4 tests in place of 142 of them fail on line 88 Claude Codereview job Support Where's my order? It's been 70 minutes Your rider's GPShasn't moved for 25minutes. I'verefunded ₹598. orders#5521 Thali 349 + Roll 249₹598 ON THE WAY rider idle 25 min payments order 5521the customer's order order 5512another customer's ₹640 refunded verification skipped:12 of 100 chats PreToolUsehookruns before the tool runs under ₹1,000: the agentrefunds it ₹1,000 or more: a person form₹4,800 customer careSwiggy's code agent loop chatget_order, track_riderresultsissue_refundresultSwiggy's codeissue_refund ₹4,800 bot comments in one month 412 of 610 resolved without a change .github/review-prompt.md report: bugs and security problems skip: style and what the linter checks a severity scale with code examples three worked examples invoiceRangrez kajal12 at ₹655 extract_price_list "rate_inr":555655 "quantity": 12"stated_total": 17700 retry with the exact errorline items add up to 16,500, not stated_total 17,700 1 2 3 4 5 Message Batches APIhalf the price, results can take up to 24 hours4,000 requests, each with a custom_id review jobNykaa's codeNykaa's codebatch4,000 requests searchbrief document-analysisbrief synthesisbrief reportbrief 18 tools, one line each analyze_contentAnalyzes the given content and returns key information. analyze_documentAnalyzes a document and returns key information. the November bulletin,a PDF 19 of 20 to the right tool search_newshospital visits, November 2026 30 s result A[ ]reads asfound nothing result BisErrorThe news archivedid not answerwithin 30 seconds. Inshorts' codecoordinatorbriefInshorts' codebulletinInshorts' codeisError one call per file 1 2 3 4 5 6 7 8 9 10 11 12 13 14 one integration callhow the files connect resume--resume edd-test-plan fork fresh startsummary 70 turns case facts order 7712₹860promised 7:10 pmdelivered 7:52 pm sent fresh with eachcustomer message asks fora person handed overstraight away,before any checks a person case factsorder 7712, ₹860 document-analysisreads the bulletins AQI 352government monitorsAQI 418private sensor networkhospital visits86% correct Delhivery's codeClaude CodeSwiggy's codeSwiggy's codeInshorts' codecoordinator Agentic Architecture & Orchestration27%chapters 7, 8, 12, 15 and 16Tool Design & MCP Integration18%chapters 1, 5, 13 and 14Claude Code Configuration & Workflows20%chapters 2, 3, 4 and 6Prompt Engineering & Structured Output20%chapters 9, 10 and 11Context Management & Reliability15%chapters 17 and 18

The chapters in order

Swiggy, Inshorts, Delhivery, Razorpay, PhysicsWallah and Nykaa appear in this course only to explain how Claude Code and Claude work. None of these companies is connected with these tutorials, and the code, systems, numbers and events in the chapters were written for the lessons. They do not describe how these companies work, and the companies do not use this code.