Skip to main content

Agent Tools, Skills, And Export

Some Document.Bot features are more advanced than basic search and chat, but they are still user-visible.

Agent Tools​

The AI agent can use tools to work with your workspace. Depending on your build and settings, tools can include:

  • reading files
  • reading document tables of contents
  • searching documents
  • querying similar content
  • converting documents
  • preparing file edits
  • creating draft outputs

Tool activity should be visible enough that you can understand what the agent is doing. Sensitive file actions should remain reviewable.

Human Approval For File Actions​

When an agent wants to write or change files, review the proposed action before accepting it.

Check:

  • file path
  • source evidence
  • proposed content
  • whether the file is inside the intended workspace
  • whether the task should create a draft instead of editing the original

Skills​

Skills are reusable instructions or workflows that can guide the AI for a specific kind of task. Depending on your build, you may be able to browse, create, edit, enable, or disable skills.

Use skills when a task needs repeatable rules, such as a preferred review checklist or report format.

In chat, type $ to mention an enabled skill when you want to deliberately use that workflow. Type @ for files or folders instead.

Some builds include official skills from Document.Bot. General skills may be enabled by default, while specialized skills can start disabled so they do not affect normal document Q&A. Enable a specialized skill only when the task matches that workflow.

Disabled skills stay visible in the Skills page but are not offered to the AI agent at runtime. This helps keep ordinary chat, summarization, drafting, and explanation prompts from accidentally triggering specialized review behavior.

Do not enable skills you do not understand for sensitive workspaces.

Official Skills And Support Files​

Some builds include official Document.Bot skills for repeatable document workflows, such as audits, decision briefs, source-pack preflights, evidence packages, document-control hygiene checks, traceability matrices, and verification or validation reviews.

Official skills can include read-only support files, such as templates, reference notes, checklists, or example registers. These files help explain how the workflow should run. They are not ordinary workspace evidence, and they should not be treated as proof about your documents.

Specialized official skills may start disabled by default. Enable one when the task clearly matches that workflow. Leave it disabled for ordinary Q&A, summaries, rewriting, brainstorming, or broad explanations.

When a chat step says it is using a skill, you may be able to open the skill viewer and inspect its instructions or support files. This helps you understand why the agent followed a particular checklist or output format.

MCP Servers​

MCP server settings can make external tools available to the AI. This is powerful and should be treated carefully.

Before enabling an MCP server:

  • understand what it can access
  • check whether it can read or write files
  • confirm whether it can use the network
  • disable tools you do not need

Some builds can let the agent use internet search. Use it when current or external information matters, but keep workspace evidence separate from web evidence.

For document review, ask the agent to label which claims came from local files and which came from the web.

Export​

Document.Bot can support export workflows such as turning Markdown outputs into PDF or HTML when available.

Use export after you have reviewed the source-backed content. Export is a publishing step, not a substitute for checking the draft.

Diagnostics​

Beta builds may include diagnostics such as performance trace recording or debug context for feedback. Diagnostics are useful when reporting slow startup, indexing problems, rendering issues, or agent behavior problems.

Do not attach private files or secrets to diagnostics unless you intentionally prepare a safe reproduction package.

Last updated on