Word & Excel add-ins

After installation: which Skill to pick, what each setting does, and which combinations match real-world tasks.

Connect an add-in to an agent

Both add-ins need to know which agent to call. The pattern is identical:

1. In Knowledge → Agents → pick or create the agent you want to use.
2. Open the agent → Keys tab → + Mint key. Give it a friendly name ("Word — Sarah's laptop"). Copy the token shown once.
3. In Word or Excel: open the add-in task pane → Add a connection → paste the API base URL (https://ask.olyteck.com) and the token → Save.

The token is stored per-device by the add-in (Office's encrypted settings store) and remembered across sessions. Each key is bound to a single agent — mint separate keys for separate agents and switch between them in the add-in's Connections tab.

Skills — what the add-ins actually do

Once you're connected, the add-in shows a Skill dropdown. Each skill is a pre-built workflow that bundles three decisions for you: how to call the agent, what to read from the document, and where to put the output. Pick the skill that matches the task; the defaults are right 90 % of the time.

Word — six skills

Excel — two skills

"Use as context" — what the agent reads

For Word, this controls what gets sent alongside your prompt as document context. Most skills set a sensible default, but you can override it.

• None (prompt only) — the agent uses only its KB and your prompt. Use when you're generating new content from scratch and the surrounding document isn't relevant.
• Selection — only the text you've selected goes as context. Use for targeted operations on a paragraph or sentence (translate, rewrite, fact-check just this bit).
• Current section — the paragraphs under the heading where your cursor is. Use for section-scoped operations (rewrite this section in our house style).
• Whole document — every paragraph, numbered for citation. Use when the task needs full context (summarize, verify, audit).
• Headings only — just the document outline. Use when you want the agent to know the structure but not be biased by existing body content (fill-template skill uses this).

\'Output\' — where the result lands

How the add-in inserts what the agent returned. Different skills default to different modes; override when needed.

• Insert at cursor — straightforward append. Use for drafting, summarizing, generating new content. The most common choice.
• Replace selection — overwrites your highlighted text. Use for translate, rewrite, polish. Always work on a copy of the document first — there's no undo specifically for add-in writes (Office's standard Ctrl+Z works but only one step at a time).
• As Word comments (audit) — every finding becomes a Word comment anchored to the paragraph it's about. Use for any review-style skill (Verify / Audit). Reviewers see the comments in the side pane just like teammate review comments.
• As tracked changes — the agent's edits show as suggested redlines that the author accepts or rejects one-by-one. Use when you want machine-suggested edits with a human final say (regulated documents, legal contracts).
• Appended report table — a structured table at the end of the document listing each finding. Use when you want a printable audit deliverable rather than inline comments.

Choosing combinations — real-world patterns

"I want to draft a new section from scratch"

Skill Free prompt · Context None · Output Insert at cursor.

Type the prompt ("Draft the methodology section for a managed-services contract, scope: ITSM, deliverables: standard runbook + 24/7 monitoring"). The agent draws from the KB for examples of past methodologies and writes the section. Insert it where your cursor sits.

"I want to rewrite this paragraph in our house style"

Skill Free prompt · Context Selection · Output Replace selection.

Highlight the paragraph. Prompt: "Rewrite in the formal voice of our standard {document type}." The KB provides house-style examples; the rewritten version replaces your selection.

"I want to fact-check this policy before publishing"

Skill Verify document · everything else auto-set.

The agent extracts every checkable claim, verifies each against the KB (e.g. your existing standards, prior policies, certifications). Problematic ones land as Word comments with: the verdict (contradicted / not in KB / partially supported), the source it conflicts with, and a suggested KB-aligned rephrasal. Reviewer accepts, rejects, or rewrites each.

"I'm reviewing a long policy doc — does each section comply with our standards?"

Skill Audit document · everything else auto-set.

The agent loops through every heading + body block, checks compliance against KB rules, flags missing / incomplete / non-compliant items as comments. Long-doc-safe: each section is its own LLM call, so you don't hit context limits.

"I've laid out a doc skeleton with empty headings — fill in the body"

Skill Fill template from sources · everything else auto-set.

Headings serve as the prompt structure. The agent generates body content under each one, grounded in the KB. Best used when your headings are descriptive enough that the agent can infer intent ("3.2 Backup retention" vs the cryptic "3.2 Other"). Edit the output afterwards — this gets you a first draft, not a final document.

"Translate this paragraph to French / German / Ukrainian"

Skill Translate selection · the prompt field is the target language.

Highlight the source text, type "French" in the prompt, run. The translation replaces the selection, preserving paragraph breaks and inline formatting.

"Summarize this 60-page report into one page for the board"

Skill Summarize document · place cursor where you want the summary.

The whole document is sent as context. Output appears at the cursor. Tip: drop a heading like "Executive summary" first, then trigger the skill on the blank line below.

The "Add source citations as footnotes" option

Available on most Word skills. When ticked, every claim the agent generates gets a Word footnote with the source document name and page number. Use for anything customer- or auditor-facing — citations are what differentiate a grounded answer from a credible-sounding hallucination.

Excel — the row-by-row killer use case

This is what Knowledge does that no generic chatbot does well. You paste a column of questions, you get back columns of answers — with confidence and citations — at speed.

The basic flow

1. Open a spreadsheet. Put each question in column A, one per row. A header row is fine; pick the input range to skip it.
2. Open the Knowledge task pane (Home ribbon → Knowledge).
3. Pick the connection (which agent to use) and the skill (Answer questionnaire or Free prompt).
4. Pick the input range — usually A2:A200 or similar.
5. Pick the output column — usually B. If the connected agent is in JSON mode, the output spreads across multiple columns (B, C, D, E…) — one per JSON top-level field.
6. Click Run. Rows process one at a time; you watch them populate live.

JSON mode + column mapping

If your agent's output mode is JSON with a schema like:

{
  "answer":     "string",
  "confidence": "number",
  "sources":    ["string"],
  "gaps":       "string"
}

then the add-in maps each top-level field to a column:

This is what makes Knowledge different from copy-pasting into ChatGPT 200 times: the file comes back ready to send to the prospect / auditor, with citations they can verify.

The questionnaire review pattern

For 100-500 question deliverables (SIG, CAIQ, custom security questionnaires, RFPs):

1. Run the agent over the whole column.
2. Sort by confidence ascending. The bottom 10-20 % are the rows a human must touch.
3. Manually answer or fix the low-confidence rows.
4. The other 80 % is already publishable. Save and ship.

This converts a 5-day RFP into ~1.5 days end-to-end and a 3-week security questionnaire into ~2 days.

Things to watch out for

• Run on a copy of the workbook. The add-in writes into cells; an "undo" after a 200-row run is annoying.
• Stop and inspect after the first 5-10 rows. Cheaper than discovering at row 180 that the system prompt or schema was wrong.
• Per-key rate limits. If you configured a rate limit on the agent key (see Agents), Excel respects it. The bar shows pause/resume.
• The agent ignores cells outside the input range. If you change the input range mid-run, those new rows aren't processed until you re-run.

Agents that pair well with Excel

• RFP Responder — JSON output, low temperature, scoped to bid library project.
• Security Questionnaire Responder — defensive system prompt ("abstain if not in evidence"), grounding mode on, JSON output.
• Audit Evidence Mapper — each row = one auditor control, output = document name + page + evidence summary.
• Contract Clause Comparator — each row = one clause, output = deviation + severity + recommended response.

Calling the same agent from your own code

The add-ins are clients over a stable JSON API. If you want to wire Knowledge into a procurement workflow, a CI job, or a custom internal tool, the endpoint is:

POST /api/run.php
Authorization: Bearer <your-agent-key>
Content-Type: application/json

{ "input": "Do you encrypt data at rest?" }

Returns the same JSON shape the add-ins consume. No SDK required — every modern language can do this in 5 lines.