Free tools Windows power users keep installed
One-click scans. No signup required.
iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
AI agents can add useful context to a shared workspace, but their notes should not silently become trusted reference material. A safer pattern separates a human-curated source library from temporary agent-written notes: agents can read approved sources and append observations, while a person checks each note before keeping it or promoting its claims.
That distinction is the core of Deian’s human-in-the-loop knowledge-base workflow. The author describes building it into Takibi Base after agents confidently reused outdated Notion material. The account is one operator’s experience, not evidence that all Notion workspaces or agent setups fail this way.
Why separate sources from agent notes?
A shared knowledge base has two different jobs: preserve material people have chosen to trust, and capture temporary observations that may be useful later. Combining those jobs makes it easy for an agent’s unverified output to look as authoritative as a reviewed source.
Deian’s design draws a clear boundary. The source library contains durable documents the author selects and organizes. Agent notes are scratch work: explicitly untrusted, append-only, and not promoted automatically. The author puts it simply: “A note does not become a fact by being saved.”
This addresses a familiar maintenance failure in the author’s own workflow: agents could write to Notion, but did not reliably update existing material. The resulting notes could be stale or contradictory while still appearing current. The lesson is not that agents should never write; it is that writing permission is not the same as authority to revise shared knowledge.
How the workflow handles a note
Agents read approved sources and append observations
In the described system, retrieval gives an agent short verbatim passages from the source library, along with openable citations and a support score. This makes the evidence behind an answer inspectable instead of asking a person to trust an unsupported summary.
Agents can append notes, but cannot edit them in place. Notes do not cross Profiles, do not become sources automatically, and expire after the author’s chosen 30-day period. That expiry is a workflow setting, not a duration validated by a study or applicable to every deployment.
People decide what survives
Each note reaches a human triage queue with three possible outcomes:
Rank #2
- Keep: retain the note as a note rather than treating it as an approved source.
- Add to draft: move relevant material into a draft, where the person checks and edits its claims before anything enters the library.
- Remove: discard material that is stale, unsupported, irrelevant, or otherwise not worth keeping.
The human review is the promotion gate: agents can contribute, but a person decides whether a claim belongs in durable reference material. Deian summarizes the rule as: “The rules that matter: agents append, never edit. Nothing graduates without a human reading it.”
What should happen when retrieval finds too little?
A knowledge system should make an evidence gap visible rather than fill it with confident-sounding prose. In the described workflow, a retrieval with no match returns no passages, and the agent is instructed to acknowledge that it lacks supporting material. When sources conflict, the system surfaces the competing passages together so a person can judge them.
This behavior is useful only if users can inspect the underlying evidence. A citation should open the relevant source, and retrieved text should remain distinguishable from the agent’s interpretation. If the library is thin, queries may need to be broader, followed by search or reading the full document; retrieval cannot provide evidence the library does not contain.
Limit each agent job to the access it needs
Separate trust levels are stronger when access controls enforce them. Deian describes Profiles that expose defined views of the workspace, so an agent job can receive access to the material and actions it needs without receiving broad editing rights.
For example, a transcript uploader can add files to an Uploads area but cannot edit or delete the source library. Uploads permission is off by default in other Profiles. Keys can be rotated or revoked individually, and pausing a Profile stops the keys associated with it. These controls limit the consequences of a mistake or compromised credential; they do not replace human review of a note’s claims.
What the operator has to maintain
Human review is deliberate work, not a free by-product of adding an agent. Deian calls the triage queue a real bottleneck. The queue also does not rank older notes differently, so the operator still has to decide what deserves attention. A small library can make retrieval less useful and require more searching or full-document reading.
The author reports a library of 87 documents, but that is a self-reported size, not a verified benchmark or a recommended target. The useful question is not how many documents to collect; it is whether the set is curated, relevant to the agent’s tasks, and maintained well enough that a citation can support the answer.
Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Repair Windows errors before they cause bigger problems3Fix the driver behind crashes, sound loss and screen glitchesA lightweight version using folders
A dedicated product is not necessary to adopt the central rule. The author proposes a basic alternative: keep approved sources and temporary notes in separate folders, use a 30-day expiry habit for notes, and review the notes regularly. The 30-day interval is the author’s chosen setting, not a universal retention standard.
- Create two distinct locations. Put human-approved reference documents in one folder and agent-generated notes in another. Give agents read access to sources and append access only to the notes location where possible.
- Make the boundary explicit. Tell agents that notes are untrusted, must not be edited in place, and cannot be treated as approved facts. Require source passages or citations for claims drawn from the library, and require an explicit acknowledgement when no supporting passage is found.
- Review notes on a regular schedule. For each note, decide whether to keep it temporarily, move checked claims into a draft for possible source inclusion, or delete it. Do not copy claims into the library until a person has checked them.
- Expire old scratch material. Delete notes past the chosen retention period, adapting any cleanup script to your actual folder paths and backup or recovery requirements. A deletion command can permanently remove files; test it on a disposable directory before using it on real notes.
This folder-based approach provides the trust boundary and human promotion step, but does not automatically supply citation-aware retrieval, conflict presentation, per-job Profiles, or credential controls. Those capabilities require additional tooling or a more managed system.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Choosing how much system to build
The right implementation depends on the consequences of a bad note and the amount of content being managed. The table compares the two approaches described here; it is not a controlled performance comparison.
| Consideration | Folders and regular triage | Takibi Base as described by its author |
|---|---|---|
| Setup and maintenance | Organize two folders, establish a retention habit, and run regular review. Cleanup behavior must be adapted to the environment. | The author says the workflow is built into the product; further current product details are not stated in the cited account. |
| Note retention and promotion | Depends on the operator enforcing expiry and the human review step. | Notes have a 30-day TTL, are append-only, and do not become sources automatically. |
| Retrieval evidence | Depends on the search or retrieval tools added to the folders; not specified as an inherent folder feature. | Described retrieval returns short verbatim passages, openable citations, and a support score; no-match and conflicting-source behavior are also described. |
| Permissions | Depends on filesystem and tool permissions configured by the operator. | Profiles scope workspace views and actions; keys can be rotated or revoked, and pausing a Profile stops its keys. |
| Review burden | The operator handles the review process and retention routine. | The author still describes triage as a bottleneck; the queue does not yet rank stale notes differently. |
Takibi Base is the author’s implementation, not a demonstrated requirement for the pattern. The source account does not establish current pricing, availability, or a comparative advantage over other tools.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
