Recently, I was getting feedback that some of the engineering tickets I prepared with Claude were too prescriptive. They gave engineers possible fixes to pick apart, rather than a clear account of the work and room to decide how best to do it.
I’ve been developing a broader Claude skill for product work on an internal project. It covers backlog and epic management, with estimation and GitHub workflows still being worked through. Ticket preparation is only one part of it. For that work, the skill can inspect code and existing tickets, draw on research and previous decisions, and prepare a ticket in the team’s format. It’s helped us challenge explanations that didn’t match the code and question assumptions before adding more work to the backlog.
In the way we’re working at the moment, the ticket is an important handover between Product and Engineering. We don’t rely on regular refinement sessions to resolve everything afterwards. The written work item needs to carry enough information for somebody using AI tooling to pick it up.
Part of the difficulty came from our own instructions. Early templates invited a root cause and suggested implementation. Later versions included standard regression and verification criteria. Some of the prescription wasn’t AI inventing work, the templates encouraged it.
So I started pushing the skill towards describing outcomes, particularly for features, without deciding the implementation on Engineering’s behalf.
What disappeared in the rewrite
In October I used AI to rewrite 18 existing tickets in a more outcome-focused form. Beforehand, I’d explicitly warned that important details must not be lost.
That instruction wasn’t enough. We kept the originals, and independent AI comparison checks subsequently identified small omissions in five of the rewritten tickets. Nine edits restored the missing information. There is no evidence that the omissions caused a delivery problem and none of the missing material was unrecoverable.
But the nature of those omissions mattered to us. They included who needed to agree a change, qualifications about existing behaviour and safeguards concerning access and limits. We considered those details worth restoring.
These were conditions or context we’d intended to carry forward, even as we removed unnecessary implementation instructions.
A ticket can be less prescriptive and less informative at the same time. In this case, telling AI not to lose important detail didn’t remove the need to check what changed.
Two questions before a ticket is ready
I’ve found it more useful to ask what each detail in a ticket represents than whether there’s too much or too little of it.
The ticket needs to explain the outcome and carry forward constraints we’ve agreed. A finding from the code might be useful for an engineer to investigate, but that doesn’t make it the confirmed cause or the fix. Suggested approaches can still be included without turning them into requirements, while anything the team hasn’t decided should remain an open question.
I’ve been building that distinction into the skill’s instructions. One check asks, for each proposed acceptance criterion: would this still have to be true if Engineering chose a different valid implementation? If not, the line may be a proposed solution rather than a condition for success. That doesn’t mean discarding the suggestion, it means putting it in the right place.
A second check compares the proposed ticket with its sources and, where there is an earlier ticket, with the original text. Did a constraint disappear? Did an open question quietly become a decision? Did a suspected cause become a verified fact? Every criterion should be traceable to the source or an actual agreement, rather than appearing simply because AI generated it.
One rewritten ticket said a technical finding had been checked against the main branch of the code. An engineer later challenged part of it, so the wording carried more certainty than the finding could support.
Something readers can use
For someone using a different ticketing setup, the useful part is checking where each acceptance criterion came from, whether it unnecessarily limits Engineering, and what changed between the source and the draft. They don’t need our particular Jira template to do that.
The prompt below takes just that narrow part of the wider skill and makes it something readers can try and adapt to their own working practices. It isn’t the whole skill or a ready-made ticket-management system. Unlike the internal workflow, it can’t access a codebase or backlog: it can only work with what you paste in. It’s a starting point rather than a proven improvement to delivery and its comparison still needs human scrutiny.
To try it, paste the original requirement or ticket alongside any constraints, exclusions and scope decisions already agreed. Otherwise, the assistant cannot tell which details should survive the rewrite.
You are helping me turn a requirement into an engineering ticket without losing
requirements or prescribing implementation. You cannot inspect my systems, code
or documents beyond what I paste here. Say so whenever a point depends on
something you cannot see, and never fill a gap with an invented design.
1. UNDERSTAND. Restate the user, problem, outcome, constraints, exclusions and
agreements in my source. Ask me about intent and outcome only, never about
mechanisms. If the source is only a title, ask me for the intent.
2. SORT FACTS. Label every statement about the current system VERIFIED (I gave
evidence), MY WORD, or UNVERIFIED (you inferred it). Keep my claims at the
strength I gave them. A comment or mock-up is a proposal, not a decision,
unless I confirm it. An open question stays a question. Put diagnoses and
fixes in a separate "hypotheses" section.
3. DRAFT. Write who needs what and why, then acceptance criteria that someone
outside this conversation could mark done or not done, one outcome each.
Test every line: would it still be true under a different, equally valid
implementation? If not, move it to "Technical notes (optional approaches)".
Add nothing I did not ask for or imply: no tests, coverage, documentation or
monitoring criteria. Where a journey can be empty, refused or fail, state
what the user sees, or list it as an open question.
4. CHECK. Show a table, one row per check, with Pass / Fail / Judgement call and
a comment, never a score: criteria testable; no mechanism in the body; every
fact labelled; nothing added unasked; failure paths covered; open questions
still questions. For each Fail, propose replacement wording for me to accept
or reject.
5. COMPARE. Check the draft against my original source and any scope decisions I
gave you during this chat. Report RETAINED / REMOVED (with reason) / ADDED
(who asked for it) / CHANGED (meaning or certainty) / UNCERTAIN. Restore
anything dropped without my agreement.
6. CONFIRM. Show each criterion with why it is written and its source, quoted.
I will answer Keep, Amend or Drop for each. "All as proposed" confirms
nothing. Re-show anything amended. Finish with a list of the decisions I made,
ready to record on the ticket.
For a higher-stakes ticket, paste only the draft, source, checklist and decision
list into a fresh chat and ask it to review them.In this prompt, “VERIFIED (I gave evidence)” means supporting material was supplied, not that the assistant independently verified the underlying claim. The source still needs to be assessed.
Consider a fictional bookseller. The initial request says customers need to export their own order history for the last seven years, without accessing another customer’s data. Whether guest checkouts are included hasn’t been decided.
An over-prescriptive AI ticket might specify an endpoint, database table, queue and library, then add a test-coverage target and treat guest checkouts as included. If I asked it to simplify the ticket, it might come back with customers can export their order history. That would lose the seven-year requirement, the access restriction and the fact that guest checkouts are still undecided.
A useful ticket keeps the seven-year range and the ownership and data-isolation rules as requirements. It leaves guest checkouts as an open Product question. The technical approach can be offered for Engineering to evaluate, without becoming an acceptance criterion. If AI adds a case such as what happens when there are no orders, that’s another question to resolve, not a new requirement to slip in unnoticed.
The comparison helps me spot where a requirement has been dropped, a suggestion has become mandatory, or an open question has been treated as decided. I still need to check the source and decide whether the proposed changes are right.
The cost of checking
This doesn’t make the review problem disappear. An assistant comparing its own rewrite with the source can still miss something. That is why the original needs to remain available for review. A person has to read the original, the draft and the comparison, particularly where scope or confidence has changed.
I don’t yet have evidence that the evolving skill saves time across Product and Engineering, or that engineers prefer the later versions. Producing a ticket quickly is only one part of the work.
Over the last few months, we’ve also been questioning whether tickets are the right way to organise work at all. There’s value in having work documented, understanding how it connects to our strategy and being able to measure whether we’ve achieved what we intended. But there’s a balance to find. The detail we add to make a ticket clear and ready for development can start to prescribe the solution, when that judgement should sit with the engineers doing the work.
We’ve been experimenting with different approaches, and I’m not sure we’ve settled on the right balance yet. I still think we need to capture why something matters, what we’re trying to achieve and which decisions or constraints have already been agreed. What I’m less certain about is how much of that needs to be written into a ticket, or whether the ticket itself is always the right way to manage it.
I still want AI to investigate and help prepare the handover. But before I approve a cleaner ticket, I want to know what it took out, what it added and what it quietly decided was true.


