Documentation Index
Fetch the complete documentation index at: https://docs.isotopes.ai/llms.txt
Use this file to discover all available pages before exploring further.
How aidnn works
aidnn is built to take a finance or business question, plan an approach to answering it, pull from the data sources you’ve connected, and run the analysis. That has a few practical implications for how you prompt:
- aidnn does its own planning. You don’t need to tell it the steps — you tell it the outcome you want, and it figures out how to get there.
- aidnn looks at your data and your prompt together. If you reference a file, sheet, or table by name, aidnn uses that. If you don’t, it makes its best guess based on what’s available.
- aidnn verifies its own work. Most analytical outputs go through a verification step before you see them. This is a deliberate design choice — it’s what makes aidnn outputs reliable enough to bring into executive conversations.
- aidnn works best within a focused session. A single notebook session is where aidnn carries context. Starting a new notebook is a fresh start.
1. Lead with the goal, not the steps
Tell aidnn the outcome you want — not the steps to get there. This is the most important rule of thumb to keep in mind.
When you over-specify the steps, you can accidentally constrain aidnn in ways that produce a worse answer than if you had simply stated what you wanted to learn.
Example
| Avoid | Recommended |
|---|
| ”Pull the billing table from Snowflake. Filter to active customers. Group by month. Sum MRR. Calculate churn. Calculate expansion. Subtract churn from expansion. Divide by starting MRR. Plot it." | "Generate NRR from Jan 2024 to July 2025 using billing data in Snowflake. Plot both MRR and NRR.” |
Rule of thumb — Tell aidnn what you want to know, what time window, and what source. Let it figure out the steps. If the answer comes back wrong, that’s when you add specificity — not before.
2. Give context about your goal
Provide aidnn with context around your question so it can orient its analysis and give more accurate, tailored results. The more it knows about what you’re trying to accomplish — and who the output is for — the better the response.
Example
| Avoid | Recommended |
|---|
| ”Analyze this spreadsheet." | "I’m preparing a quarterly vendor spend review for our leadership team. I’ve uploaded our contract tracker and spend data. Identify the top 10 vendors by total committed spend, flag any contracts expiring in the next 90 days, and highlight where actual spend exceeds committed amounts.” |
3. Be specific about data sources
aidnn does its best to figure out which data to use, but if you don’t name the source clearly, it falls back on assumptions based on file or table names. Sometimes those assumptions are right; sometimes they’re not. The fix is to be explicit.
Example
| Avoid | Recommended |
|---|
| ”Summarize Q3 OPEX by cost center." | "Using the FY24_Actuals sheet in the Budget_Master file, summarize OPEX by cost center for Q3.” |
If a workbook has multiple tabs and you don’t name one, aidnn may pick the wrong tab — or pull from a tab labeled “raw” when you wanted the cleaned version. Reference the specific sheet name when there’s any chance of ambiguity.
How aidnn interprets naming conventions
aidnn uses naming patterns to infer what a table is for:
- A _mapping table (e.g.,
dept_to_exec_mapping) — aidnn assumes it maps values from one field to another and joins it automatically.
- An _enriched table — aidnn assumes the mapping has already been applied and treats it as cleaned, ready-to-analyze data.
- A _raw or _source table — aidnn treats this as untransformed input and may try to enrich it before analyzing.
If your data doesn’t follow these conventions, that’s fine — but reference the specific file, sheet, table, or column by name so aidnn doesn’t have to guess.
4. Specify quantities and boundaries
Tell aidnn exactly how many items, rows, or time windows you need. This avoids over- or under-reporting and reduces follow-up corrections.
Example
| Avoid | Recommended |
|---|
| ”Show me contracts that are expiring soon." | "Show contracts expiring within three time windows: 0–30 days, 30–60 days, and 60–90 days from today. Exclude contracts that have already expired. Group by vendor category, sorted from highest to lowest committed spend.” |
5. The rules are easy — the exceptions are where detail matters
The core logic of most analyses is straightforward — sum this, divide by that, compare year over year. aidnn handles the standard rules well. Where users get tripped up is the exceptions: the company-specific quirks of how your business actually books, categorizes, or reports things.
Examples of exceptions aidnn won’t know unless you tell it
- How your company defines “active” customers (paid in last 30 days? Has a contract? Has logged in?)
- Which GL accounts to include or exclude in OPEX (e.g., do you exclude one-time restructuring charges?)
- What counts as “large” — top 10 by spend? Anything over $1M? Anything strategic regardless of dollar size?
- Whether intercompany eliminations have already been applied to the data you’re uploading
- How fiscal periods map to calendar periods (especially for 4-4-5 calendars)
- Whether bookings are gross or net of discounts in your reporting
Example
| Avoid | Recommended |
|---|
| ”Add savings calculations to the summary." | "Calculate estimated savings using these rules: If Status = ‘Deprecate’: savings = 100% of committed spend. If Status = ‘Renew’ and Scope = ‘Decrease’: savings = 30%. If Status = ‘Renew’ and Scope = ‘Renegotiate’: savings = 15%. Show a summary row at the bottom with total savings opportunity.” |
Keep a short list of your team’s standard exceptions (e.g., “exclude account 7420 from OPEX,” “only count customers with a signed MSA”) and paste the relevant ones into your prompts. After a few uses, you’ll have a personal library of prompt fragments you can reuse.
6. Structure complex requests step by step
When your request involves multiple steps — pulling from different tabs, cross-referencing files, applying business rules — break it down into a clear, numbered sequence rather than writing one long paragraph.
Example
| Avoid | Recommended |
|---|
| ”Take the tracker template tab and make it a dashboard. Pull links from the contract export tab, pull department and spend from the other file, build it so I can export a CSV and leave blank columns for status and strategy." | "Build a contract tracker from the uploaded files: 1. Start from the ‘Tracker Template’ tab as the base layout. 2. From the ‘Contract Export’ tab, pull: Vendor Name, Links, Expiration Date. 3. From the ‘Spend Report’ file, pull: Department, Annual Spend, Category. 4. Leave these columns blank for manual entry: Status, Strategy, Action Required. 5. Output as an Excel file.” |
7. Building dashboards and spreadsheets
Dashboards and structured spreadsheets are a common aidnn use case. A few specific tips:
Be explicit about the structure you want
If you want a single consolidated output, say so. If you want multiple deliverables, call that out specifically. aidnn handles “give me one consolidated workbook with these tabs” better than open-ended requests.
“Build a single Excel workbook with two tabs: Tab 1 is a master tracker showing all contracts with vendor, value, start date, end date, and owner. Tab 2 is an expiration dashboard showing only contracts expiring in the next 90 days, sorted by expiration date.”
Think in views, not just tables
Dashboards work best when each tab or view has a clear purpose. Instead of “build a dashboard,” try:
“Build a dashboard with three views: a summary view showing totals by category, a detail view showing the full record list sorted by value, and an exception view flagging anything expiring within 30 days or exceeding a spend threshold.”
Specify sort order
If you want results sorted a particular way — highest spend first, soonest expiration first, alphabetical — say so explicitly. “Sort by [column] descending” is the clearest phrasing.
Bundle formatting instructions together rather than spreading them across follow-ups. aidnn handles formatting most reliably when stated clearly up front.
“Format dollar columns as currency with no decimals. Bold the totals row. Highlight any contract expiring in the next 30 days.”
If you’re new to aidnn, start with the structural and analytical content first, then layer in formatting. Get the right data into the right shape, then refine the visual presentation.
Automate your branding — If you find yourself repeating brand colors, fonts, or formatting guidelines in every prompt, ask your account contact to set up a Branding SME for your account. Once configured, aidnn automatically applies your brand theme — colors, header styles, chart palettes — to every artifact it produces across all users on your account. No prompting required. After branding is set up, you never need to say “use our brand colors” again. If something comes back off-brand, that’s a bug to flag to your Isotopes AI account team, not a prompting issue.
8. One ask per turn after the first
Your first prompt can (and should) describe the full outcome. After that, iterate with one focused request per turn. Combining unrelated requests in a single follow-up leads to messier, less verifiable output.
Example
| Avoid | Recommended |
|---|
| ”Build the tracker, export it, schedule it weekly, and write a team SOP — all in one go." | "1. Build my weekly contract tracker from the uploaded files. 2. Schedule this to run every Monday at 8 a.m. 3. Generate an SOP for my team on how to use and maintain this tracker.” |
9. Iterate — don’t start over
When you’re close but need adjustments, make targeted follow-up requests instead of asking aidnn to “redo” everything. Small, specific tweaks are faster and preserve the work already done.
Example
| Avoid | Recommended |
|---|
| ”Redo the whole dashboard." | "Two adjustments to the dashboard: 1. Add a ‘Renegotiate’ option to the ‘If Renewing’ dropdown, with an estimated 15% savings. 2. Filter out the Travel & Entertainment category from all tabs.” |
10. When aidnn gets it wrong
- One-line correction first. “The West region should map to Lisa, not James — please re-run.” Cheapest, fastest. aidnn fixes and continues.
- Say “Verify it.” aidnn runs verification on most outputs automatically, but you can also explicitly invoke it as a follow-up. Especially useful after adjustments or before sharing.
- Ask it to show its work. If a number looks off, ask: “The report shows 320 contracts due in 0–30 days — that seems high. Walk me through how you calculated that.” Other useful follow-ups: “What data did you use?”
- Flag platform issues. If something feels like a bug — files not loading, formatting broken, scheduled runs not firing — flag it to your account contact. Those are platform issues, not prompting issues.
Common pitfalls
| Pitfall | Fix |
|---|
| Bundling many unrelated asks into one prompt | State the outcome first; iterate one ask per turn |
| Saying “dashboard” when you mean “spreadsheet” | Be specific: “Excel workbook with a summary tab and detail tabs” |
| Treating aidnn like a search engine (“OPEX Q3”) | Use full-sentence prompts with the goal, time window, and source |
| Not naming the sheet in a multi-tab workbook | Always reference the specific sheet by name |
| Re-uploading the same file in different versions | Remove or label superseded files to avoid using the older version |
| Assuming context carries across sessions | Each new notebook is a fresh start. Restate key definitions |
| Not specifying the output format | Say “chart,” “table,” “Excel,” “PDF,” or “paragraph” explicitly |
| Repeating brand colors in every prompt | Set up a Branding SME — then it’s automatic |
| Vague “redo” without saying what changed | ”Two adjustments: [1]… [2]…” |
Quick-reference prompt patterns
| If you want to… | Try a prompt like… |
|---|
| Run a metric over time | ”Calculate [metric] from [start date] to [end date] using [data source]. Plot it monthly.” |
| Compare actuals to budget | ”Compare [period] actuals to budget for [scope]. Show variance in dollars and percent. Flag anything over [threshold].” |
| Test multiple scenarios | ”Model the impact of [variable] at [list of values]. Show side by side.” |
| Build a dashboard | ”Build a dashboard with [N] views: a summary of [metric A], a detail view of [metric B] sorted by [column], and an exception view flagging [condition].” |
| Build a tracker | ”Build a single Excel workbook with [N] tabs: [describe each tab’s purpose, columns, and sort order].” |
| Get a written explanation | ”Summarize the key drivers of [metric change] in plain language for a non-finance audience.” |
| Verify an answer | ”Verify it.” — works as a follow-up to most analytical outputs. |
| Format an Excel export | ”Export to Excel. Format dollar columns as currency. Bold the totals row. Use [color] for header rows.” |
Frequently asked questions
How specific does my prompt need to be?
Specific enough that someone reading it cold would know what you want, but not so specific that you’re listing every step. State the goal, the time window, and the source. Add company-specific exceptions when they matter. Skip the rest.
Why does aidnn sometimes take longer than I expect?
aidnn runs a verification step on most analytical outputs. This adds time compared to systems that don’t verify, but it’s what makes aidnn outputs reliable enough to bring into executive conversations.
Can aidnn pull from my data warehouse, or do I need to upload files?
Both. aidnn supports direct connections to data warehouses as well as uploaded files (Excel, CSV). For one-off analyses, uploads are often fastest. For recurring workflows, a direct connection is cleaner.
Does aidnn remember things across sessions?
Within a notebook session, yes — aidnn carries the context of what you’ve discussed and what data you’ve referenced. Across sessions, no — each new notebook is a fresh start. Restate key definitions and exceptions when you open a new one.
What’s the best way to get better at prompting?
Practice with real questions you’d ask a colleague. When something doesn’t work, look at the prompt — was the goal clear? Was the source named? Were the exceptions called out? Most prompting issues trace back to ambiguity in one of those three areas.
Questions or feedback?
Reach out to your Isotopes AI account team. We’re actively gathering feedback from users, and the more we hear about how aidnn fits into your workflows, the better we can shape the product to support them.
