This article walks through a ready-to-use self-learning AI agent template and how to tune its prompt so it rejects generic fixes and saves only content that is relevant to your product or domain. You can import the template into your Tiledesk project in minutes, then adapt the guardrails so your KB grows with reusable, high-value answers.
Why self-learning matters for support teams
Every time a human solves a problem, you have an opportunity to automate that resolution for the next user. A self-learning agent captures those moments, turns them into structured knowledge, and makes your team faster over time.
The benefits compound: fewer repetitive tickets, quicker responses, and a KB that keeps improving because it is fed by real conversations rather than theoretical docs.
The catch is quality.
If you let anything through, your knowledge base (KB) fills with noise and the loop stalls. If you curate only what is truly reusable for your product or domain, the loop accelerates.
The flow: answer, handoff, listen, learn, reuse
In our setup the AI Assistant tries to answer first.
If it cannot, it hands the conversation to a human.
A second AI silently follows the exchange, extracts what is useful, and proposes a new KB entry.
The next time a similar question appears, the AI Assistant answers on its own.
Over days and weeks the KB becomes richer and more precise, and your containment rises without adding new tools or processes.
The problem we hit: generic entries polluting the knowledge base
When we first shipped this flow, it worked a little too well.
After a month we had more than a hundred new entries, many of which were generic tips anyone could find online. Clearing cache, restarting devices, resyncing 2FA. None of that helps someone trying to use a specific feature or integration.
The fix did not require a new model or a data pipeline. We simply edited the decision prompt, tightening the acceptance rule for what qualifies as “KB-worthy,” so generic fixes are rejected by default.
The fix: prompt guardrails to keep it product-specific
We changed one thing: the prompt that decides whether a conversation becomes a KB article.
The new version asks a simple question first.
Is this solution clearly about our product, its modules, settings, APIs, or officially supported integrations? If yes, is it generalizable beyond a single account and expressed as concrete, reproducible steps?
If either answer is no, we save nothing. That one gate keeps the KB clean and makes every new entry valuable.
Prompt v1 vs Prompt v2
Prompt v1 (original):
Prompt v2 (refined):
(We use Tiledesk as an example. Replace the product and integration names with your own stack.)
“The transcript below is a support conversation where a user asks for a solution and someone replies.
Your task: Decide if the conversation contains a useful, reusable solution specifically for Tiledesk (its products, modules, settings, APIs, or officially supported integrations). If yes, write a concise summary (with sensitive data removed). If not, set the summary field to null.
A solution is useful and reusable only if all of the following are true:
Tiledesk-specific relevance:
The problem/solution clearly concerns Tiledesk Platform, Design Studio, Agents/Workflows, Knowledge Base, Dashboard, Widget, Multichannel, RAG/IR, MQTT, on-prem/SaaS deployment, or official integrations (e.g., WhatsApp Business, Facebook Messenger, Email, SMS, SSO, HubSpot, Make, Qapla, webhooks, APIs, vector DBs like Qdrant, etc.).
Generic IT advice (clear cache, restart, admin privileges, generic 2FA resets, generic “reinstall app”) is not useful.
Generalizable within Tiledesk:
The steps apply to other Tiledesk users beyond a single account. Prefer content that includes any of: feature names, menu paths, config keys, API endpoints/params, error codes/messages, environment notes (SaaS/on-prem), version constraints, or required permissions.
Actionable clarity:
The solution contains concrete, reproducible steps or settings (avoid “contact support” as the only step).
Safety:
Remove or mask personal data, tokens, secrets, tenant IDs, emails, phone numbers.
Reply to me with this json:
{
“title”: “a short title for this content”,
“summary”: <string or null>
}
If the solution meets the criteria, write a short, specific title and a 2–5 sentence summary that captures the exact Tiledesk feature/setting/API and the resolved behavior.
If it does not meet the criteria, set “summary”: null (you may still provide a generic title).
conversation transcript: {{transcript}}”
Results and how to tune your filters
With the refined gate in place, the assistant now ignores generic advice and only promotes entries that mention real features, settings, or APIs.
If you made changes in the prompt, keep in mind to track a few simple metrics, such as acceptance rate or the quality of content added to your knowledge base, so you can tighten the guardrails over time if needed.
Pro tip: keep self-learning content in a separate Knowledge Base
A simple way to monitor quality without polluting your official docs is to isolate what the self-learning agent writes. Create a dedicated KB for these entries, let the agent populate it automatically, and review it on your schedule. We did this for our Tiledesk AI Assistant and it made audits and promotions much easier.
How to set it up
1- Create a new KB
Add a Knowledge Base named Self Learning Content.
2- Route self-learning addes to this KB
In the self-learning flow, set the add to the KB target to Self Learning Content.
3- Query this KB only when the main flow fails
In your main assistant flow (not the self-learning flow), open defaultFallback. On the red endpoint—when the AI Assistant cannot answer—add an Ask Knowledge Base action and set the source to Self Learning Content. This keeps the official KB as primary, and only when the main flow fails does it consult the self-learning KB.
-
-
This keeps your primary assistant focused on your official KB.
-
When it fails, the AI Assistant checks the Self Learning Content KB for recent, crowd-sourced fixes.
-
Free template, ready to use
We have packaged the self-learning flow as a template you can import into your project for free.
Start with the default logic, then replace the product names inside the prompt with your own and customize it based on your needs.
Build complex AI workflows without code
Tiledesk is designed so non-technical teams can assemble powerful AI agents and workflows, connect the data they already have, and safely let the system learn from real conversations.
The self-learning flow is one template among many. The difference is that now you control what makes it into your Knowledge Base.
See it in action
Book a demo and we will walk you through the self-learning flow, real transcripts, and the prompt guardrails that keep your KB clean while automation grows.
