A thirty-second text edit. Two hours to ship it.
Change one sentence in a prompt and an AI product behaves differently. Yet that text usually lives in a document - no version history, no way to test a change before it ships, and no way to undo one afterwards. Imprompto is our own product, and it fixes that.
- What it is
- Our own product
- Function
- AI product operations
- Users
- AI product teams and agencies
- Stage
- Open beta, live
The behaviour left the code, and nothing followed it
For thirty years, what a software product did lived in its code. Changing it meant an engineer changing logic - and every step around that had decades of tooling built for it: version control, review, staging, rollback, an audit trail.
In an AI product, most of that behaviour has moved into the prompt: a paragraph of ordinary English telling the model who it is, what it may say, and what to refuse. That is not logic. It is text, and it gets treated like text.
So the most behaviour-critical asset in the product ended up as the only one with no infrastructure around it.
Four things break at once
The edit takes thirty seconds. Everything around it takes an afternoon - and the four failures compound.
Nobody can say which version is live. The document says one thing, the codebase another, and whoever last changed it is on holiday. Every conversation starts with an archaeology exercise.
Changing behaviour needs a deployment. Hours of release process for a reworded sentence. The real cost is not the deployment - it is that small improvements stop being worth making.
Testing happens after release. Comparing properly means running old and new against the same inputs, on the models you actually use. By hand that is tedious enough that almost nobody does - so the comparison happens in production, and the test subjects are customers.
There is no diff, so there is no way back. When behaviour degrades the first question is always "what changed?" Without history it is unanswerable, and recovery becomes a rewrite from memory.
The question is not "should we buy a prompt tool?"
It is this: how long does it take your organisation to change what its AI does, who is allowed to do it, and can you prove afterwards what it did?
Every organisation running an AI product already has an answer to those three questions. Most have never been asked them.
What we built
Five stages. Nothing reaches a customer because somebody saved a draft.
- 01
Written by whoever owns the behaviour
An editor with live preview, built for a product manager or a support lead rather than someone comfortable opening a repository.
- 02
Tested before it becomes a change
The same prompt against several models at once, on real calls, with cost and latency per run.
- 03
Every save becomes a version
Immutable, with an author and a note. Any two compare line by line; any one restores in a click.
- 04
Going live is a separate, recorded decision
A version reaches customers only when someone activates it, and that activation carries a name and a time.
- 05
The product asks for the live version
Your application requests the active prompt as it runs. That one connection takes the deployment out of the loop.
Where a prompt is written
On the left is what the author edits; on the right is what a customer receives. Variables are recognised as they are typed.
- Editorthe template
You are a support agent for {{company}}.
Reply in a {{tone}} tone, under {{max_words}} words.
Never promise a delivery date you cannot verify from the order record.
- Variables
- 4 detected
- Previewwith the variables filled in
You are a support agent for Northwind.
Reply in a friendly tone, under 80 words.
Never promise a delivery date you cannot verify from the order record.
- Defaults
- saved automatically
The one technical idea here - that a prompt is a sequence of messages, not one block of text - is shown as blocks you reorder, not a structure you learn.
Interface shown with representative data, not real customer records.
"After a long time I've seen a product with UI this clean - smooth transitions, clicks, everything feels polished. Reminds me of Google-level product quality. The workspace design is really well thought out."
One prompt, three models, the same instant
Before this existed, the product could tell you what your prompts were - not whether a change had made them better.
You are a support agent for {{company}}. Reply in a {{tone}} tone, under 80 words. “Where is my order?”
- GPT-4oChosenOpenAI
Hi! I've checked your order and it's at the distribution centre. It should arrive by Thursday. Anything else I can help with?
- Latency
- 0.8s
- Cost
- $0.0021
- Claude Sonnet 4Anthropic
Hi there! I can see exactly what's happening with your order - it's out for delivery Thursday. Shall I text updates?
- Latency
- 1.1s
- Cost
- $0.0034
- Gemini 2.5 ProGoogle
Hello! Your order is currently in transit and is scheduled to arrive Thursday. Let me know if you'd like tracking updates.
- Latency
- 0.6s
- Cost
- $0.0008
Real calls, on the team's own vendor accounts - so the cost and latency are theirs, not a benchmark from somebody else's prompts. Saving the winner promotes it to a version, so the comparison ends in a decision.
What actually changed
When behaviour degrades, the first question is always the same one. This is the answer, in seconds.
v2.4
- You are a support agent for {{company}}.
Removed: Reply in a strictly professional tone.Removed: Do not use emojis under any circumstances.- Keep responses concise.
v3.2 · live
- You are a support agent for {{company}}.
- Added: Reply in a friendly, professional tone.
- Added: You may use at most one emoji per response.
- Added: Never promise a delivery date you cannot verify.
- Keep responses concise.
Version history
- v3.2live2h ago · Priya S.warmer tone, allow one emoji - promoted from the Playground
- v3.11d ago · Priya S.shorten reply limit
- v3.03d ago · Arjun M.
- v2.46d ago · Arjun M.imported
- v2.311d ago · Arjun M.
A team that can undo a change makes more changes, and a team that makes more changes improves faster.
Going live is its own decision
Newest and live are deliberately different things - collapsing them would mean anyone editing a draft is editing production.
Activation record · Customer Support - Order Status
- 2h agov3.2 activatedservingPriya S.“warmer tone, allow one emoji - tested against Claude and Gemini”
- 1d agov3.1 activatedPriya S.“shorten reply limit”
- 3d agov3.0 activatedArjun M.“initial production version”
This is the whole deployment. Keeping saved and live apart is what lets someone rework a live prompt at four in the afternoon without a customer noticing.
Who owns what
Not that AI teams need fewer engineers - that two groups of people have been sharing one workflow which suits neither.
Behaviour - the product owner
- Write and revise the prompt
- Test it across models
- Read the outputs and judge them
- Choose the version that ships
- Activate it - and roll it back
System - engineering
- Connect the application once
- Caching, retries and failure handling
- What the product does with the output
- Latency, cost and reliability budgets
One library, and an owner for everything in it
One prompt becomes forty inside eighteen months. The ones actually serving customers are marked, so nobody has to ask.
Prompts · Support AI Team workspace
| Prompt | Category | Collection | Updated | State |
|---|---|---|---|---|
| Customer Support - Order Status | Q&A | Support Bots | 2h | live |
| Refund Eligibility Check | Analysis | Support Bots | 1d | |
| Onboarding Email Writer | Writing | Marketing | 3d | |
| SQL Query Generator | Code generation | Internal | 5d | live |
| Escalation Classifier | Analysis | Support Bots | 1w |
A prompt stops being knowledge that one person happens to carry.
Who is allowed to change what
Roles are enforced, not advisory - a viewer cannot change production behaviour by accident.
Roles · Support AI Team workspace
| Permission | Owner | Admin | Editor | Viewer |
|---|---|---|---|---|
| View prompts | Yes | Yes | Yes | Yes |
| Create and edit prompts | Yes | Yes | Yes | No |
| Activate a version - go live | Yes | Yes | No | No |
| Invite and manage members | Yes | Yes | No | No |
| Billing and workspace deletion | Yes | No | No | No |
For an agency, each client is a separate workspace, and a client invited as a viewer sees their own work and nothing else.
What a team lead opens in the morning
One workspace, over a week. The activity feed is the audit trail itself, not a summary of it.
Support AI Team · one workspace · last 7 days
- 42
- Active prompts
- 137
- Versions
- 318
- Test runs
- 24,806
- Prompt requests
TESTING, BY PROVIDER
- OpenAI52%
- Anthropic31%
- Google17%
MOST TESTED PROMPTS
- Order Status84
- Escalation61
- Refund check47
- Onboarding38
One customer workspace, shown with representative data. These are not Imprompto's own figures - ours are further down, and they are labelled.
The safeguards that make teams adopt it
A product holding an organisation's production behaviour has to be trustworthy before it is clever.
Editing is not deploying. Drafts and experiments never touch what customers see.
Every change is attributed. The trail is automatic - nothing to remember to fill in, which is the only kind that survives a busy quarter.
Nothing is deleted. A workspace that lapses goes read-only rather than lost. Your data is never leverage.
Your vendor accounts stay yours. Test runs go through your own keys, at your own rates, with nothing to unwind if you leave.
What users told us
Practitioners across product, engineering, research and academia. Two features exist because of these conversations. Quoted with permission.
"I have dozens of prompts across different tools and no single place to manage them. Imprompto solves exactly this. It's a genuine problem for anyone who uses AI seriously, and this is the first platform I've seen built specifically for it."
Rajat GoyalAssistant Professor, UPES Dehradun "The Activity Log is very useful - you can see which prompt was changed, who changed it, and when. Quick Search makes navigating across hundreds of prompts efficient. For teams managing prompts, this is a real single source of truth."
Kunal BairwaAI Founder "The platform is very intuitive - I'd rate it 8 out of 10 for anyone getting into prompt engineering. The version history and diff view are exactly what's been missing. This is genuinely useful."
Mario (kleosr)Cursor Discord Ambassador, Italy "Prompt sharing is really good, especially being able to see the metrics - how many viewed, copied, and unique visitors. Earlier we needed a third-party platform to track that."
Vipul KumaonProduct @ Attero "This is genuinely useful - otherwise you end up searching through chat history every time. Version tracking is what makes it different from just saving prompts in Notion."
Akshit BansalAI Professional
Where it is now
Our own product figures, current as of the open beta and counted from live records, not modelled. The dashboard further up this page is one customer's workspace - none of its numbers are ours.
- Open-beta users
- 290+
- Active design partners
- 3
- Build stages shipped
- 6
Where else this applies
The product manages AI prompts. The pattern underneath does not - author, test before release, version everything, release deliberately, then watch it. It fits pricing rules, underwriting policies and compliance language unchanged.
Business-critical behaviour living outside the codebase - in documents, spreadsheets, or people's heads.
A deployment required to change something that is not really engineering.
The people who understand a process separated from the people who can change it.
No answer to "what changed, who changed it, and can we undo it?"
Multiple clients on shared systems needing real isolation and a real audit trail.
Why this sits among the client work
It is not a client engagement, and listing it as one would be misleading. It is here because it is the receipt behind a claim we make on the homepage: that we monitor what we ship.
We could not sell an operating discipline for AI behaviour without running one ourselves. Imprompto is that discipline, built as a product - so you do not have to take the claim on trust.
Have an operational problem worth solving?
Bring us the process that's expensive, slow, manual or difficult to scale. We'll spend 30 minutes understanding it and telling you - honestly - whether AI or automation can meaningfully improve it.