Tell the agent to track every question it cannot answer. The audit file becomes a queue of real gaps you can review later — coached entirely from the context file.
3 things to walk away with
Coach via context
A few lines in the context file turn the agent into an auditor of its own failures.
Write to a file
The "audit file" pattern — name it, declare its purpose, point at it from context — works for any side effect.
Failure is signal
An agent that knows when it failed is more useful than one that always answers confidently.
Unanswered Question Tracking — coach the agent to log its own failures
Add one rule to the context file: "any time someone asks a question that you cannot answer, append it to unansweredquestions.md." The next session reads that rule. When a user asks something the agent cannot answer — like "who is the legal decider of vendor terminations" — the agent attempts an answer, admits it cannot fully resolve, and writes the question into the file.
A single line of context turns the agent into a self-auditor. You did not write code, you wrote a rule — and the behavior changed for every session afterward.
The point is not the specific use case. It is the shape: any side-effect behavior you want the agent to perform (log this, route that, save here) can be specified in a few sentences of markdown and it just works.
Tracking unanswered questions requires a custom plugin or SharePoint extension. It requires a sentence in the context file pointing the agent at a markdown file to append to.
The Audit File Pattern — name it, declare it, point the agent at it
The pattern is simple and reusable: (1) create a markdown file with a descriptive name, (2) add a line to the context file that tells the agent what to write into it and when, (3) tell the agent to create the file (or do it yourself). From then on, the file accumulates entries automatically.
In the demo, the user adds one sentence to context, then says "create this file" in chat. The agent creates unansweredquestions.md in the agent assets library. From that point forward, every session that hits an unanswerable question appends to it — with the date, the question, and what the agent looked at.
The audit-file pattern generalizes. Want to log every report sent? Every decision made? Every override? Same recipe — name a file, declare the rule, the agent writes.
The agent will create the file silently the first time it needs to write. It will not — declare the file's existence (or create it explicitly) so the agent knows where to append.
Auto-Log on Failure — the agent calls out what it didn't know
When the agent encounters a question it cannot resolve, it does three things in one response: (1) attempts an answer with whatever partial information it has, (2) calls out the limitation in the chat ("I'm going to log this as an unanswered question since I wasn't fully able to resolve"), (3) writes the entry to the audit file.
The chat-side narration matters. The user sees the agent reasoning about its own failure in real time — that builds trust and lets them correct it on the spot if they have the answer.
An agent that names its own failures in chat is more trustworthy than one that always sounds confident. Visibility of "I don't know" is a feature, not a bug.
Logging happens silently in the background. The agent narrates it in the same response — the user sees the failure, the log, and the partial answer together.
Flashcards — Applied
02 · Applied
Unanswered Question Tracking
tap to reveal →
A context-file rule that makes the agent append every unresolvable question to an audit markdown file. Turns the agent into a self-auditor without code — just a sentence of markdown.
← tap to flip back
02 · Applied
Audit File Pattern
tap to reveal →
A reusable recipe: name a markdown file, declare its purpose in the context file, tell the agent when to append. Generalizes to any side effect — logs, decisions, overrides.
← tap to flip back
02 · Applied
Auto-Log on Failure
tap to reveal →
The agent's behavior when it cannot resolve a question: attempt a partial answer, narrate the failure in chat, and append the question to the audit file — all in one response.