Project retrospectives: turning good observations into a change in how work is done
Run a project retrospective that leads to a specific change: examine events, choose an owner and review the improvement in later work.
Article language: English

In this article
After a project everyone knows that next time they must “communicate better”. After the next project they know it even better. A retrospective should interrupt this cycle of good intentions: reconstruct specific events, choose a change and test it in the next piece of work. Colourful notes may help, but they cannot change a process by themselves. They need a person, a task and a moment for acceptance.
Here a retrospective means a conversation about how work was performed and the improvement that follows. This workshop is our own proposal for a service project. It does not assume that the company uses Scrum and does not replace product acceptance, incident investigation or an assessment of an individual. The examples are hypothetical and illustrate decisions.
Establish what the meeting should change
Name the scope: a particular stage, release or collaboration between roles. Do not try to solve an entire organisational culture in one meeting. Participants should know which events are covered and which elements they can influence. If an agreement needs a decision outside the team, identify the route for passing it on.
Separate assessment of the outcome from assessment of the way work was done. A website may have met requirements even though release required an unplanned overnight intervention. A project may also have run calmly without solving the recipient's problem. Both observations matter but lead to different questions. Website acceptance (in Polish) does not replace discussing how the team reached that result.
Prepare an event trail before judging
Gather important dates, material versions, approval points and examples of handovers. You do not need a complete archive of every message. Select items connected to the problem. If a claim concerns a delay, check when the material was ready and what decision it awaited. Participants' memory alone may omit dependencies that were obvious at the time.
A project decision register (in Polish) preserves reasons for choices. Use it to distinguish an execution error from a knowingly accepted constraint. Do not judge an old decision as though the team knew the later outcome. Check what was known then and which condition was supposed to trigger reassessment.
Begin with an identifiable observation
“Communication was poor” might mean no owner, a late comment, an unclear version name or a contradictory decision. Ask for an event: what was expected, what happened and what effect followed. The discussion then has material to work with. An adjective alone does not tell you which instruction needs changing.
Record what helped too. If a brief preview before release found a missing item, keep that step. A retrospective need not catalogue disappointments alone. Decide what to repeat, what to change and what is no longer needed. Otherwise the next project may accidentally remove a working practice along with an unnecessary procedure.
Distinguish inspiration from Scrum from this workshop
The November 2020 Scrum Guide describes the Sprint Retrospective as planning ways to improve quality and effectiveness. The team considers people, collaboration, processes and tools, and addresses important improvements as soon as possible. That describes a particular approach. Our project review draws on the idea but does not mean the company follows all of Scrum.
Do not copy event names without their meaning. Showing completed work to recipients and a team conversation about its way of working have different purposes. They can be connected, but allow room for an uncomfortable observation without simultaneously running a sales presentation. Choose a form suited to the problem, people and ability to make a later change.
Make room for a different perspective
Do not begin with the judgement of the most senior person. Participants may then adjust their memories to the interpretation already given. Allow a moment to write observations independently, then compare events. The facilitator should ask about conditions and effects instead of immediately deciding who is right.
In a small team, do not promise full anonymity if you cannot provide it. Explain use of notes, access and boundaries of the discussion. If an issue needs a separate procedure, do not try to solve it with a single board vote. A process-improvement workshop does not give the facilitator authority to resolve every kind of problem.
Treat the first conversation after discovering a current error (in Polish) separately. It may need prompt fact-finding and protection of the recipient of the work; do not postpone those actions until the next retrospective. A later review can then explain what needs to change in the way work is done.
Find the mechanism before choosing a remedy
As a working example, suppose website content returned for correction several times immediately before release. Possible causes include an undecided offer, review of an old version or no named approver. Each leads to a different improvement. A general “send earlier” will not fix the source if nobody knows who must decide.
Compare several similar events and look for a shared condition. Do not assume the first reason noticed explains everything. Record alternative explanations and evidence needed to distinguish them (in Polish). In AI-supported processes, principles for responsible agent workflows (in Polish) are useful: task, access and approval point deserve separate checks.
Choose a small number of changes you will actually make
A list of twenty recommendations can look ambitious and change none of the next tasks. Assess effect, effort and ability to check. Choose one or a few steps that fit the team's actual work. The number of recommendations is not a measure of meeting quality. More important is whether participants can name the next change and its owner.

Atlassian's retrospective instructions recommend assigning owners and deadlines to actions and including them in the normal work list. We use only that part of the practice, not a promise of a particular improvement. An agreement needs a place where the team actually manages tasks, instead of waiting for someone to reopen meeting notes.
Record an improvement as behaviour to test
Turn “improve communication” into a specific step. Working example: before handing over a translation, the lead identifies the approved source version and the scope of the latest change. The recipient confirms they are using that basis. It is an observable action, not a declaration that everyone will be more attentive from now on.
- Event: what problem recurred?
- Change: what will we do differently in the next similar case?
- Owner: who will prepare and run the trial?
- Place: where will an instruction or tool change?
- Assessment: how will we recognise the step's usefulness?
- Review: when will we decide whether to keep it?
Do not automatically add a new form for every difficulty. The right change may be removing a field, combining two approvals or identifying one source. Procedures cost time too. If one creates more work than the problem it solves, it should be able to fail its own usefulness test.
Account for the previous change before adding another
At the start of the next review, check what happened to the earlier agreement. Did it have an owner, available time and a place in the plan? If nobody did it, do not immediately recommend greater commitment. Identify the obstacle. The change may have required a decision outside the team or conflicted with an existing instruction. That is a concrete dependency to resolve, not proof of everyone's bad faith.
Separate non-execution from lack of effectiveness. An improvement that was not applied has not yet been tested. One applied without a comparable situation may still await evaluation. Preserve that state in the notes. The team should neither discard an idea merely because it had no chance to use it nor announce success just because a new procedure was published.
If the task keeps losing to current work, reduce its scope or deliberately abandon it. One smaller change may be more sensible than maintaining five untouched recommendations. Keep the reason with the original aim. The next person can then see a priority decision rather than a mysterious list of open items long pretending to be an active plan.
Check whether the fix moves the problem elsewhere
A new step can make one team's work easier while increasing another's load. If adding an approval, establish its purpose, owner and availability. Observe the whole task. Faster document handover does not yet shorten service if the document later waits in another queue omitted from the review.
Invite people affected by the change to assess it, even briefly on a specific case. They need not attend every detail of the retrospective. They should understand the new task and be able to raise a constraint. The shared result matters more than a local speed record. Otherwise improvement resembles tidying your desk by moving everything onto your neighbour's.
Record a condition for withdrawing the fix too. If it creates unnecessary waiting or does not reduce the identified problem, the team should be able to return to a simpler path. Not every procedure deserves permanent residence after one successful presentation. A retrospective can end by removing a step if observation supports it. The decision should remain checkable and understandable.
Practise a release retrospective
At a hypothetical company the site was deployed correctly, but the team checked the form on an earlier preview. Public acceptance revealed a different sending configuration. The retrospective does not end with “test more carefully”. Participants identify a specific moment and the missing connection between the check and the released version.
The change is to record the public address and deployment identifier with the test result. The person accepting confirms the message in the destination inbox. At the next release the team checks whether the step happened and whether it can reconstruct the evidence. This does not guarantee no future errors; it removes a recognised ambiguity in one acceptance method.
Return to the improvement at the next opportunity
Review both execution and usefulness. Was the instruction changed? Did anyone use it? Did the step help in the situation for which it was created? If no comparable case has occurred, wait to assess effectiveness. Record that the change awaits the right test rather than adding success to a completed meeting.
An improvement can be described as a business experiment (in Polish). Set an expected observation, assessment condition and ability to withdraw. If the team understands “done” differently, start with a company glossary (in Polish). Another status will not help if everyone gives it a different meaning.
Accept the change using a knowledge-work example
As an exercise, suppose a new employee finds two contradictory instructions. The team assigns one source owner and adds a link from the older document. When onboarding the next employee, it checks whether a person without knowledge of the history reaches the right version. This is a separate usability test, not merely a check that someone added a link.
If they still choose the old instruction, return to the name and place where they search. Do not end the discussion by saying the fix was made. Performing an action and solving the problem are different outcomes. A retrospective should make that difference visible without starting another project to improve everything. Sometimes a small adjustment to the same step and another calm observation are enough.
Preserve context and remove unnecessary ceremony
Keep events, conclusions, chosen actions and the review method in the notes. Do not record every statement if it is unnecessary for the purpose. After several meetings assess the format too: can participants raise a problem, choose a feasible change and return to it? A recurring ritual without effect does not become an improvement process because the calendar is regular.
To analyse an AI-supported workflow and prepare an improvement to test, see Aeneia Audit. A plan for retiring automation (in Polish) can also help when ending an unnecessary process. A meeting conclusion matters when the next task actually looks different because of it.








