Skip to content

Lesson 6 of 6

The Write-Up: What It Showed, What Was Faked and What Comes Next

Write the one page that outlasts the demo: the answer, the evidence, the limits, what a real version would need and one recommendation, and then close the proof of concept down properly.

4 minForward Deployed Engineering

After this lesson you will be able to write up a proof of concept so that the customer can act on it, and close the work down properly.

Why the write-up matters more than the demo

The demo lasts an hour and reaches the people in the room. The decision is often made a week later, by someone who was not there, from whatever was written down. If nothing was written down, it is made from memory, and memory keeps the page that looked good and drops the grey box at the bottom.

So write for a reader who missed the demo and has five minutes.

Answer first

Start with the question from the brief, word for word, and the answer. Then the evidence in two or three sentences, with the comparison.

Do not start with what you did on Monday. The order of the week is of interest to nobody but you. A reader who stops after the first four lines should know the question, the answer, how sure it is and what you recommend.

Say what was faked, and what real would take

The second job of the write-up is to stop the proof of concept being mistaken for the product. List what was faked, and next to each item what a real version needs. This list is also the first draft of the plan for the real build, which is why it is worth doing with care.

For Brightwell the real version needs things the week never touched: the data arriving every week without anyone exporting it, a place for the team to see the list and record what happened on each call, a decision about who may see which members, and somebody to own it.

One recommendation

End with one recommendation, not a menu. The customer asked you because you have seen this kind of work before. "It depends" hands the problem back to them.

The recommendation should have a next step small enough to say yes to. For Brightwell the largest unknown left is whether a call changes anything, and that can be tested before any software is built: produce the list by hand for eight weeks, call half the members on it, and compare how many of each half cancel.

The write-up

# Proof of concept write-up: Brightwell Leisure

## Question
From the visit records we already have, can we make a weekly list
of up to 15 members who are much more likely than most to cancel
in the next four weeks?

## Answer
Yes. We made the list as it would have been on 30 March and
looked at the four weeks after. Of 13 members on the list, 9
cancelled. The pass test was 1 in 3. A list picked by chance
would have held about 2.

## How sure
One back-test, one month, 106 members. The exact share would
differ in another month. The gap over chance is large enough
that the answer would not.

## What it missed
5 of the 14 members who cancelled were not on the list. 2 were
visiting less, but not by half. 3 kept visiting until they left.

## What this does not show
Whether a phone call keeps a member. Nobody was called.

## Faked, and what a real version needs
One export made by hand   -> a weekly feed from the turnstiles
A list in a file          -> a page the team opens, with a place
                             to record each call
Run by an engineer        -> scheduled, with an owner at Brightwell
Member numbers only       -> a decision on who may see names

## Data
Six faults found and handled. See the data note.

## Recommendation
Do not build yet. For eight weeks, produce the list by hand each
Monday and call every second member on it. Compare cancellations
between those called and those not. If calling helps, build the
weekly feed and the page. We can run the eight weeks with one
day of our time a week.

## Decision needed
Yes or no to the eight-week test, from the head of membership.

Notice what the recommendation does. A proof of concept that passes usually leads to "build it". Here the honest reading is that the list works and its value is still unproven, so the next step is a second, cheaper test. Recommending less than the customer expected to buy is often what earns the trust for the larger project.

Close it down

A proof of concept leaves things behind. Clear them up in the same week.

  • Hand over the code, the data note and the write-up in one folder, with a short file saying how to run each script.
  • Delete the export from your machine, and from anywhere else it was copied, on the date agreed. Tell the customer when it is done.
  • Remove any access you were given for the week.
  • Say in writing what the code is: a proof of concept, with six tests, not something to run in production as it stands.

The last point protects the customer and you. Code that worked in a demo has a way of being run every Monday for two years.

What you practised

Look back over the week. On day 0 you turned a wish into one question with a pass test. On day 1 you found six faults in the data before they could surprise you. On days 2 and 3 you built the thinnest version that worked, and tested the one rule that mattered. On day 4 you measured it against what really happened, and against chance. On day 5 you showed the customer their own data and wrote down what it meant.

None of the code was difficult. The skill an FDE brings to a proof of concept is in what surrounds it: choosing one question, agreeing the test first, being exact about what is real, and reporting the result as it is.

Key takeaways

  • The write-up is what the customer keeps; write it for someone who was not at the demo.
  • Answer the brief's question in the first lines, with the number and the comparison.
  • List what was faked and what a real version would need, so that nobody mistakes the proof of concept for the product.
  • Make one recommendation with a next step small enough to agree.
  • Hand over the code and the data note, and delete the customer's data as agreed.

Try it

Suppose the back-test had come out differently: 3 of the 13 members on the list cancelled, where chance would give about 2. Write the "Answer" and "Recommendation" sections of the write-up for that result. A good result says plainly that the pass test was not met, gives both numbers, does not blame the data or ask for more time without a reason, and recommends either stopping or one specific further test with its own pass mark.

Official documentation

Ready to use this in a job?

Rolebound applies to matching roles for you.

Start now

Written with AI and checked automatically. Nobody has run these steps by hand, and tools change, so a command may behave differently in the version you install. Spotted a mistake? Tell us.

    Cookies on Rolebound

    We use necessary cookies to sign you in and keep the site working — those cannot be turned off. We would also like to set analytics and advertising cookies so we can see which pages help people get interviews and which adverts brought them here. Nothing optional is set unless you say yes, and you can change your mind at any time from “Cookie settings” in the footer. Read our Privacy Policy.