Lantide Data
This translation is available, but has not yet been checked against the latest Traditional Chinese source.

← Back to Learning Center

Plan → annotate → Execute → Report

Read time: ~9 minutes · Series: Analyst practice · Previous: Quick vs Project · Next: SQL-first or walk through USER_GUIDE §2

Full flow at a glance

Take the question → explore schema → draft Plan → you annotate and revise → you press Execute
  → Agent runs SQL step by step → Report (numbers + limitations) → optional HTML / follow-up Plan

Mapping to the USER_GUIDE five-stage table: you control the metrics in stages 2–3, authorize the data pull in stage 4, and accept the deliverable in stage 5.

On a dual-monitor or wide-screen setup, you can use Compare view for a read-only side-by-side during the Execute and Report acceptance stages (see below and USER_GUIDE §11.10).

Stages 1–2: take the question and explore

What you do:

  • In the AI panel, state clearly in natural language: the analysis goal, the time range, and your intuition about the denominator (if you have one)
  • If the Agent follows up with interactive questions, fill in the gaps before moving on
  • Confirm the Agent checks the table structure first, rather than guessing column names out of thin air

What the Agent does:

  • Look at available tables, grain, join keys, and obvious quality issues
  • Start writing a Plan draft (not yet Executed)

Status bar: Project Focused or Planning.

Analyst takeaway: this stage can be fast, but the Plan should already start showing stage definitions, the denominator, and checkpoints—not just "let's analyze the funnel."

Stage 3: review the Plan and leave annotations (before Execute)

In the Plan, you should at least see:

  • The analysis goal
  • Each stage / metric definition
  • The denominator and the grain
  • Validation checkpoints (e.g., the biggest drop-off stage, NULL checks)

When the metrics are wrong:

  1. In Visual mode, select a section → annotate
  2. Or explain in the conversation, but important agreements are still best captured in a Plan annotation
  3. Multiple annotations can be handled at once with Resolve (N) (N is the number of open annotations in the sidecar)

The Agent only changes the annotated section, avoiding a full rewrite that also breaks sections that were fine. After Resolve, you can use View changes to compare; handled annotations can be Archived to clear the panel. If a large edit to the body causes the anchor to become outdated, you'll need to Re-anchor by re-highlighting the original text.

Iron rule: don't press Execute during the Planning stage. The Agent also won't pretend in the conversation that you already approved and sneak in a formal data pull. Executing before the metrics are settled means pushing disputes off until the Report blows up. When a project has no Plan file yet, the Agent can use add_plan in Planning to create a new Plan.

Which files did the Agent touch this round? (the file-changes bar)

When the Agent runs patch_plan per your annotations, builds SQL tabs in sequence during Execute, or produces a Report, you might have to expand several Tool Calls to see what changed. At the end of every AI reply, if that turn wrote to a persistent SQL, Plan, Report, or HTML Report, a file-changes bar appears at the bottom of the message: one file per row, with green +N / red -N line counts; click a filename to open the corresponding tab.

Found a bad patch, or a broken SQL step during Execute? Click Undo on that reply's row to revert that file all at once to its state before this reply started (multiple edits to the same file within one reply are merged into one count). If you've manually edited the body after the Agent wrote to it, the undo will be blocked—use a manual revert or ask the Agent to fix it instead. For the full limitations (a newly created file is emptied rather than deleted, add_report and Plan Executed, etc.), see USER_GUIDE §12.14.

Stage 4: Execute—only you press it

After reviewing the Plan, click Approve & Execute in the Plan toolbar:

  • Plan → Executing (body becomes read-only)
  • The system automatically notifies the Agent to start (you don't have to type "start execution")
  • The Agent asks foreground / hybrid (only these two; background agent analysis is offered in Quick Analysis only; asked just once)
  • The stages listed under Pipeline Stages always create SQL tabs; in hybrid mode, the remaining steps mostly use run_query in the conversation
  • The Progress panel shows the Todo and query steps; a red dot signals when there's an update

During execution you can:

  • Watch the SQL tabs (foreground mode)
  • After an interruption, use Continue to resume (no need to go back to Planning and start over)
  • Use the file-changes bar at the bottom of the AI reply to check which tabs and Reports were created / modified this round, and Undo a single reply's writes if needed (see above)

If the data is unavailable or the definition or decision premise has changed, use Stop & Replan… rather than marking an unfinished Plan complete. Record the reason and partial outcome; the old Plan becomes Stopped and keeps its Progress and formal evidence. You can draft a linked replacement Plan immediately or later with New Plan. The replacement starts in Planning and must be reviewed and executed again.

The Agent cannot press Execute for you; and once a Plan enters Executing it can no longer be edited, to guarantee that the "contract sign-off" is always reliable.

If a SQL query is too long or too complex, the Agent may be asked to break it into smaller steps—so you can review it step by step, not for form's sake.

During Execute, the main window is often switching between SQL tabs and Progress, while the metrics live in the Plan. A typical approach:

  1. Right-click the Plan → Open in compare view → New compare view (or pick an already-open Compare view 1).
  2. Drag the Compare view to a second monitor (or the other half of the screen); keep SQL / Agent in the main window.
  3. Inside the Compare view the Plan is read-only; if the main window still has the same Plan tab open, edits mirror in real time to the Compare view (before saving, an Info hint appears next to the compare tab).

You can also open a persistent SQL tab into the same Compare view, side by side with the Plan (up to 5 tabs), to compare "what the contract says" with "how the SQL is actually written." Temporary SQL tabs cannot be opened this way. Persistent SQL in the Compare view follows the dark / light theme; when there are query results or errors, the height of the SQL and results areas can be split by dragging.

Note: Compare view cannot Run SQL and cannot edit the Plan; all edits stay in the main window. A Reference tab is marked with an amber md badge in the Compare view, consistent with Plan (blue) and Report (green).

Stage 5: Report acceptance

After Execute finishes, the project will contain a Report (usually paired one-to-one with the Plan).

A qualified funnel-type Report should have:

  • Each stage's definition and denominator explanation
  • Concrete numbers (counts / rates), not just "drop-off is severe at some stage"
  • limitations—what the data can't answer, metric assumptions, correlation ≠ causation

The Report can be annotated and revised, and exported as an HTML Report for sharing (for external weekly meetings, where the other side doesn't need Lantide installed). To let a colleague annotate the Plan and continue the same project inside Lantide, use Export project… to package a .lantide (sharing the analysis context and results; without SQL tabs or local data)—see USER_GUIDE §11.1.0. To hand off the entire workspace (including SQL tabs and connection settings), use the Workspace export. To dig into the next question, just create a new Plan in the same project and iterate.

Analysis Lineage: Open it from a Plan or Report toolbar to view the Plan → Markdown Report → HTML Report chain. It distinguishes formal evidence from historical context, reports incomplete links, and lets you open each artifact—without treating unsaved edits or ambiguous manual links as verified evidence.

Use Compare view to accept the Report vs the Plan

When accepting, you often need to check: whether the Report's numbers cover the checkpoints in the Plan, and whether the stage definitions are consistent.

  1. Open the Report in the main window to keep annotating or fine-tuning. If the Report is already paired with a Plan, you can click Linked plan: … in View options to open the source Plan directly; if not paired, use Link source plan to pick an unpaired Plan in the same project.
  2. When you need a side-by-side, right-click the same project's Plan (in the sidebar or an already-open tab) → Open in compare view, side by side with the Report.
  3. Switch compare tabs inside the Compare view to line up the metric sections and the number tables.
  4. If the Plan on disk was updated externally and the Compare view shows an older version, an Info hint appears next to the tab; click Refresh to reload from disk.

An already-Executed Plan will also show Linked report: … in View options, or offer Link generated report when not yet paired. Manual pairing only adds the Plan / Report association; it does not mark the Plan as Executed, freeze its steps, or complete the Todo; a formal run still has to go through Approve & Execute.

If an external Agent collaborated on this round

The acceptance standard doesn't change because of the conversation interface: you still verify the Plan, formal SQL evidence, and the Report and limitations. Additionally open External MCP Activity to confirm which artifacts were actually modified, and that the +N/-N and open actions match expectations; a single content change that can be safely rolled back offers Undo, and if the file has since been modified it is blocked rather than force-overwritten.

An external connection's mode is a capability ceiling, not the same as the Plan being approved. Even when using Execute or Admin, a formal analysis still has to be pressed by a human via Approve & Execute. For connection choices, Activity, and how to revert, see External Agent Integration; for the full operations, see USER_GUIDE §13.7.

Common analyst questions

Q: Can I skip the Plan and just tell the Agent to run SQL?
In Quick you can run exploratory queries; in a Project, a formal deliverable should go through Plan → Execute, otherwise you'll lack a contract when the weekly meeting asks follow-ups.

Q: I found the Plan was wrong after Execute?
While it is running and its premise changes, use Stop & Replan to preserve partial evidence and draft a replacement Plan. After Executed, the Plan body is locked; a major change usually means a new Plan. Minor fixes to the Report can use annotations (depending on state, see USER_GUIDE §11.6).

Q: Will the Agent run statistical tests during Planning?
The Planning stage focuses on exploration and revising the Plan; formal statistics happen after Execute, and the Agent will confirm parameters with you first (see statistical analysis tools).

Q: Can I edit annotations inside Compare view?
No. Annotation cards are for reading and hover comparison only; View changes lets you review the edits of a Resolved annotation. To add / edit annotations or Re-anchor, go back to the main window.

Q: How do I view an annotation with an outdated anchor inside Compare view?
The Compare view shows the card and metadata for an outdated anchor (equivalent to Anchor outdated in the main window), but cannot Re-anchor—handle it back in the main window.

Q: If I close a SQL tab in the main window, can I still see it in Compare view?
Yes. The Compare view keeps a content snapshot from the moment of closing and can still be compared read-only; when you need the latest results, click Refresh in the Compare view.

Next steps

← Previous: Quick Analysis vs Project AnalysisNext: SQL-first: metrics written in tabs →