Case study: what we got wrong and fixed while building our own content system.

The situation

In the last few weeks we built a large set of guides and small tools for Indian founders, coaches and creators. We wrote them with help from AI tools and checked them with people. That is the same way we work with clients, so we treated our own site as the opening project to get right.

This case study uses just things we can show you: our own lessons log, our own checklist and our own mistakes. It does not contain a client result, a sales figure or a story about a named customer. We do not have permission to share those, and we will not invent them.

Why write it at all? Because a team that just shows wins teaches nobody anything. The useful part of any process is what it catches. So here is what ours caught.

The scene

It is Sunday morning. Twenty four new guides went live that same morning, and the team is pleased. Then a voice note arrives from Madhur. Paraphrased, it says: these are too short, write them so a person stops and finds them useful, like the earlier big article.

He was right. The pages worked, but they were thin, and no checklist had flagged it. That moment is why this case study exists. It is about the gap between work that passes a checklist and work a reader keeps.

What other case studies do, and what we did differently

Before rewriting this page we looked at how case studies in our space are usually written. We read search results and page summaries, not full reports, so treat this as a quick scan and not a study. Our notes are labelled snippet level for that reason.

• Many open with one big result, such as money recovered or a percent lift. Those figures come from the seller of the service, and we could not verify any of them.
• Most follow the same arc: problem, solution, result. It is clear, but the result carries the whole story.
• Few show a mistake. A reader cannot learn how a team checks its work when just the wins appear.
• Few give the reader something to copy. You finish the page admiring the work, not able to repeat it.

So we made three choices. We lead with what went wrong, because that is where the learning is. We state the rule behind each fix, so the reader can reuse it. And we use no result figure, because the sole honest figures we have are about our process, not about clients.

That is a smaller claim than most case studies make. We think it is a stronger one, because every line on this page can be checked against our own log.

What we set up before we started

We follow a simple loop we call the Taste Test: draft, check, fix, learn. The person who checks is not the person who made the piece. For bigger pieces we add four reviewers, each looking through one lens.

Each lens gives a score out of ten. Anything under nine goes back to Fix. We also keep a lessons log, one dated line per mistake, so the next draft starts better than the last.

Mistake 1: pictures that did not belong

We made six new articles and reused crops of images from other articles to save time. A reviewer noticed that readers would see the same faces and scenes again. We regenerated fresh images for each piece.

A smaller version of the same problem came a day later. Some pictures showed a corner shop, a print shop and a sweet shop because the writer drew the picture from the comparison in the text. The reader was an Indian founder in an office, not a shop owner. We replaced seven images and wrote the rule down: the comparison lives in the words, and the picture shows the reader.

Mistake 2: two numbers that disagreed

A small calculator on one page showed a rate as 4.5 per 1,000 in one place and 0.4 percent in another. The two views rounded the same number in different ways. Nobody saw it while typing round values, and a reviewer found it when testing a value that landed on a half.

Readers trust a tool less the moment two numbers on it disagree, even when both are close.

Mistake 3: a quote with no source in the opening lines

One article paraphrased a public figure in its quick answer without naming where we read it. The quick answer is the part AI answer tools lift on its own. If it travels without its source, the claim travels without its evidence.

Mistake 4: a deadline line that did not match the fact

A share line said to send the page before December, while the dated rule it referred to started near the end of December. It sounded urgent and it was not accurate.

We care about this one because false urgency is easy to write and hard to forgive.

Mistake 5: a source we thought was dead

One source gave us an error when we fetched it with a simple request, and we almost marked it as unusable. Opened in a real browser, it loaded fine. The opening check was too simple.

Related: some publisher pages block full text but give title, authors and date through a public record. When that is all we can verify, we cite just that, and we say it comes from a secondary source.

Mistake 6: small build bugs that readers would have felt

• A new tool reused a style name that another tool on the page already used, so the page threw an error. New tools now use their own names.
• Numbered lists inside boxes touched the box edge on a 390 pixel phone screen. We added padding and now check one boxed list in the phone view.
• Wide data tables got crushed on a phone. We switched to stacked cards.
• A page folder that grew past a size limit failed to publish twice. We now keep preview folders small from the start.

None of these is exciting. All of them would have made a reader feel the page was not finished.

Mistake 7, this week: the articles were too thin

This one came from the founder, not from a checklist. We had published 24 new guides in one day. Each worked, but they were short, around 700 to 1,000 words, next to our longer pieces. Madhur said plainly that they were too small, and that readers should be able to study a page like a case study later.

We agreed, said so, and rewrote them. Each now has worked examples, a step plan, a mistakes list, a checklist and more questions answered. They run from about 1,460 to 2,400 words by our count. We also told him what was still short, rather than saying it was done.

Notice what this mistake needed: someone willing to say it out loud. A checklist did not catch it. A person did.

What we measure, and what we do not claim

We do not publish a figure for how much better the pages got, because we do not have a verified one. What we measure is process: how many checks failed, what each failure was, and whether the same mistake came back.

The most useful number in this whole project is a count: how often the same mistake repeats. If a rule works, that count stays at zero.

What you can copy

You do not need a team of four. You need a habit.

1. Keep a lessons log. One line per mistake: date, what, fix, rule.
2. Have someone else check your work before a customer sees it.
3. Test every tool with an awkward value, not a nice round one.
4. Name the source of every claim in the same sentence.
5. Look at your work on a small phone screen.
6. Read your own pictures: do they show your reader or your analogy?
7. Say out loud when something is too thin, and fix it.

If you run just one of these this week, run the opening one. A log turns mistakes into rules.

A 7 day plan to start your own loop

1. Day 1: pick one thing you publish every week.
2. Day 2: write a five line checklist for it.
3. Day 3: ask one person to check your next piece against the list.
4. Day 4: fix what they find.
5. Day 5: write one line in your lessons log.
6. Day 6: repeat on a second piece.
7. Day 7: read your log and update the checklist.

What we would still change

Our process is not finished. We want a second reader on every piece before it goes live, and a real phone test on more devices. We have not done both for every page yet, and we say so. We also want to add real customer chats, with names removed and permission given, so future case studies can show customer work, not just ours.

Written by Madhur Tyagi, founder of The Human Algorithm. 8+ years, 25+ brands, since 2017.

Scope note: Scripts and plans are examples, not a promise of results. Nothing on this page is sent unless you press the WhatsApp button. No outside facts or rates are quoted.