> ## Documentation Index
> Fetch the complete documentation index at: https://gridos.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.
# Chain mode: let the AI build your workbook end-to-end
> Chain mode auto-applies every AI step, observes formula results, and continues building until your entire workbook plan is complete.
Chain mode removes the manual Apply step from the regular chat flow. Instead of reviewing and clicking Apply after each AI response, you send one prompt and GridOS loops — applying writes, reading formula results, and generating the next section — until the entire plan is done or you stop it. It's designed for large multi-section builds where clicking through every step would be tedious.
## When to use chain mode
Revenue, COGS, gross profit, OpEx, EBITDA, and a DCF terminal value all in one run — chain mode handles the full structure without you clicking Apply on each row.
Any model with more than three or four sections benefits from chain mode. The agent tracks which sections it has written and which remain.
Chain mode observes the computed values of formulas after each write, so it can use those values to inform subsequent sections — for example, using a computed revenue figure to seed the COGS row.
Describe a full workbook in one prompt and let the agent build it end-to-end, just like opening a template but tailored exactly to your spec.
## How chain mode works
Type your full build request in the chat composer and select **Chain** mode (instead of the default single-turn mode) before submitting.
```text theme={null}
Build a 4-quarter operating model: revenue starting at 100 with 10% QoQ growth,
COGS at 40% of revenue, gross profit, OpEx fixed at 20, and EBITDA
```
On the first turn, the agent emits a plan — a structured preview of every section it intends to fill, with target ranges and notes. This plan is shown to you before any cells are written. The agent also writes the **first section only** on this turn.
```text theme={null}
Plan: Quarterly Operating Model
├── Header row B2:F2 Metric, Q1–Q4
├── Revenue B3:F3 100, 10% QoQ growth
├── COGS B4:F4 =MULTIPLY(revenue, 0.4)
├── Gross Profit B5:F5 =MINUS(revenue, COGS)
├── OpEx B6:F6 Fixed 20 per quarter
└── EBITDA B7:F7 =MINUS(gross_profit, OpEx)
```
After writing a section, GridOS applies it to the kernel immediately — no click required. It then reads back the computed values of any formula cells and feeds those observations into the agent's next prompt.
Informed by the observed formula results (and any warnings), the agent writes the next section. This repeats until every section in the plan is complete.
When all sections are done, the agent signals completion by returning `values=[[""]]` — an empty write. GridOS stops the loop and presents the final summary of all steps taken.
## The plan preview
The plan is informational — it shows you the full structure the agent intends to build before cells are filled. It is emitted only on the first turn and suppressed on all subsequent turns.
Each section in the plan includes:
| Field | Description |
| ---------- | ------------------------------------------------------------ |
| **label** | Human-readable name for the section (e.g. "Revenue", "COGS") |
| **target** | The cell range the agent will write (e.g. `B3:F3`) |
| **notes** | Brief description of the formula logic or values |
The plan is a preview, not a commitment. If the agent detects that a section needs to be retried (because of a column-alignment warning), it rewrites that section before advancing — the plan section count may not match the number of loop iterations exactly.
## Column alignment auto-retry
GridOS detects the most common formula error in financial models — referencing a label cell instead of the adjacent data cell — and triggers an automatic retry when it occurs.
After applying each section, GridOS checks whether any formula references a text-valued cell. If it does, the agent receives a detailed warning and is told to re-emit the same section with corrected cell references **before moving to the next section**.
```text theme={null}
*** COLUMN ALIGNMENT WARNING ***
Formula in C4 references non-numeric cell B3 (value: "Revenue").
Re-emit the COGS section with corrected formulas. Each formula in column X
must reference cells in column X, not the label column.
```
This retry does not count as a wasted iteration — sections are only counted as cleanly written when they produce no warnings.
## Iteration limits
Chain mode runs for up to **10 iterations** by default. Each application of a section (including retries) counts as one iteration.
If your model has more than 10 sections, chain mode will stop before completing it. Break the build into two prompts — run chain mode for the first half, then run it again for the second half.
You can reduce the limit by passing `max_iterations` in the API request if you want shorter runs during development.
## Stopping chain mode
Chain mode stops automatically when any of the following happen:
* The agent returns an empty `values=[[""]]` completion signal
* The maximum number of iterations is reached
* The agent has no formula observations remaining and the plan is complete
You can also dismiss the chain at any step from the UI — the cells written so far are kept; only the remaining steps are cancelled.
## Chain mode vs. regular chat
| | Regular chat | Chain mode |
| -------------------------- | -------------------------------------------- | ---------------------------------------------------------------- |
| **Apply step** | Manual — you click Apply after each response | Automatic — applied immediately after each step |
| **Formula observation** | Not fed back to the agent | Yes — computed values are included in the next prompt |
| **Column alignment retry** | You see the error and re-ask manually | Automatic — the agent retries the section before moving on |
| **Plan tracking** | Agent tracks progress informally | Agent tracks sections written vs. remaining, with explicit hints |
| **Best for** | Single writes, quick edits, questions | Large multi-section builds, full model generation |
For large financial models, combine chain mode with [templates](/guides/templates): open a template to get the skeleton, then use chain mode to fill in the model with your specific assumptions in a single prompt.