---
title: "Preview the impact"
pattern_id: "preview-the-impact"
trait: "Trustworthy"
type: "Humorphic pattern brief for coding agents"
spec_version: 3
canonical_url: "https://humorphism.com/patterns/preview-the-impact"
markdown_url: "https://humorphism.com/patterns/preview-the-impact.md"
index: "https://humorphism.com/llms.txt"
all_patterns_in_one_file: "https://humorphism.com/llms-full.txt"
normative_section: "Rules"
rule_prefix: "PI"
rule_ids: ["PI-R1", "PI-R2", "PI-R3", "PI-R4", "PI-R5", "PI-R6", "PI-R7", "PI-R8"]
---

# Preview the impact

## How to read this file

You are an agent. Someone gave you this file so you can add one behavior to their product: the AI teammate in it should show the exact change and its reach before anything costly or irreversible runs, and keep the way back easy. This file is the complete specification for that behavior. Build from it.

**The Rules section is the only normative part.** Everything else exists to make the rules buildable: the capabilities tell you which parts of the pattern the product can support, the reference screens show you the behavior in motion, the examples give you three shapes it can take, and the checklist is how you confirm you are done. Examples and the checklist point back to rule IDs rather than restating requirements, so each requirement lives in exactly one place. If any other section seems to contradict a rule, the rule wins.

Work in this order:

1. **Find where this belongs** in the product: any send, publish, deploy, payment, delete, or bulk edit; any action that reaches many people or records at once; and the moment right after such an action runs.
2. **Read Rules.** All eight apply to whatever you build.
3. **Check Capabilities** against what the product already has. Build the examples its capabilities support, and take the stated fallback for the rest.
4. **Pick your examples.** No example can be built without a capability. If you build only one, build Example 1, which needs only Dry run.
5. **Build it in the product's own design system, components and voice.** The reference screens say what each screen does and why. They are not a specification of how it must look. Keep the behavior, change the expression.
6. **Confirm against the checklist** before you call it done.

Humorphism is a design philosophy for products with AI in them. It replaces the user interface, built for operating tools, with the human interface, built for collaborating with AI teammates. There are 21 humorphic patterns grouped into 7 traits. You do not need the others to build this one, but if you can fetch URLs: [llms.txt](https://humorphism.com/llms.txt) lists every humorphic pattern with a one-line summary and a link, and [llms-full.txt](https://humorphism.com/llms-full.txt) has all of them in one file.

## The pattern

_Trait: Trustworthy. Makes work, authority, limits, and uncertainty visible_

An AI teammate shows the draft before it hits send. Before anything costly or irreversible, the blast radius is visible and the way back is easy.

**Build this.** An AI teammate shows the exact change, lets the person edit it and try it small first, because mistakes caught in preview cost nothing.

**Avoid this.** A tool sends to 1,204 customers on one click, because it was built to execute and never to show.

The difference to carry into your build:

| Tool | AI teammate |
|---|---|
| Preview not available | The exact change, as it will land |
| One button sends to everyone | A small test first, as easy to take as the full run |
| The change is fixed once drafted | The change can be edited before it commits |
| Can’t be undone | Undo in reach, or a short delay before it commits |
| Executes, then goes quiet | Says what it did and what it will watch |

## Rules

Normative. All eight hold for everything you build from this file. Each one says what to do, why it matters, and what it looks like once it is built.

| ID | Rule | Why | Built correctly when |
|---|---|---|---|
| **PI-R1** | Show the exact change before committing anything costly or irreversible. | The person can only catch a mistake they can see. | Before such an action runs, the person sees the actual output (the message as it will land, the before and after values), not a description of it. |
| **PI-R2** | Name the reach: who or what the action will touch, and how many. | The size of a mistake is set by how far it goes. | The preview states the audience or the records affected ("All 1,204 customers") in the same view as the action. |
| **PI-R3** | When the reach is broad, offer a small test run first. | A mistake caught on 50 people costs almost nothing; on 1,204 it costs the campaign. | A broad action has a small-sample option ("Test with 50 first") that takes one click, the same as the full run. |
| **PI-R4** | Let the person edit the change inside the preview. | A preview the person can only accept or cancel turns every small fix into starting over. | Any part of the previewed change can be edited in place, and the preview shows the edit before anything commits. |
| **PI-R5** | After a test, report what happened and what the AI teammate will watch, and wait for the person before the full run. | A test nobody looks at catches nothing. | After a test, one line says what happened and what the AI teammate will watch ("I sent it to 50", "I’ll monitor performance"), and the full run waits for the person. |
| **PI-R6** | Keep a way back in reach: Undo after the action runs, or, where reversal is not possible, a short delay with Cancel before it commits. | When the way back is one click, acting stops being a gamble. | Undo stays visible for a window that fits the action, and using it restores the prior state with one line naming what was restored and to when ("The sheet is back to how it was at 2:14 pm."). Where reversal is not possible, the action waits in a visible delay with Cancel, and says it cannot be reversed after that. |
| **PI-R7** | Treat an undo as a signal about the action, not just a reversal. | An undo the AI teammate forgets invites the same mistake again. | After an undo, the AI teammate does not repeat the reversed action unasked, and within the same task it asks first before a similar action. |
| **PI-R8** | Keep the preview to one view, with commit and cancel one click each. | A preview that takes work to read gets skipped, and then it protects nothing. | The change, its reach and the actions fit in one view, and approving and canceling each take a single action. |

**PI-R1 carries the pattern.** Build the real preview, computed from the action itself, before you polish the buttons around it. A build that satisfies every rule except PI-R1 has produced a confirmation dialog: the person is asked to approve something they cannot see, which is the tool version with an extra click.

## Capabilities

Build the examples the product's capabilities support. Do not stub a capability the product lacks, and do not write copy that implies one it lacks.

| Capability | What it is | Used by | If the product lacks it |
|---|---|---|---|
| Dry run | An action’s effect computed without applying it: the diff, the records touched, the people affected. The same path can be run against a small sample first. | Examples 1 and 2 | Show the drafted change itself, the message or the new text, and skip the small test. Do not show counts of affected people or records, or a test run, that the product cannot compute. |
| Undo | Each action is recorded with enough state to reverse it, and Undo stays in view for a window that fits the action. Where reversal isn’t possible, the action sits in a short delay before it commits. | Example 3 | Hold the action in a short delay with a visible Cancel before it commits, and say it cannot be reversed after that. Do not show an Undo the product cannot honor. |

Every example needs at least one of these. **If you build only one example, build Example 1**, which needs only Dry run.

## Reference screens

These describe the animations on the pattern page frame by frame, so you can build from this file alone. They show the behavior, not the appearance to copy. **Keep the behavior, change the expression:** build in the product's own design system, components, type and voice.

One convention to translate rather than reproduce. On these screens a small orange square precedes every line the AI teammate speaks, so it is always clear who is talking, and the AI teammate's side sits on a warm, slowly moving gradient while the tool's side sits on flat gray. On the email, the AI teammate's line sits in a tinted strip under the message rather than in a chat. Before and after values carry small "Before" and "After" labels. What the person presses changes state in place: a button turns black, or turns green with a check, and a chosen row's number turns black. Use whatever your system already uses for "the AI teammate is speaking", "this is a preview", and "this is what was chosen".

### The AI teammate version (frames 01 to 03, a loop of about 9.5 seconds)

One email card, centered on a warm, slowly moving gradient of pink, orange and blue. The white sheet at the top holds two fields, "To" "All 1,204 customers" and "Subject" "Spring sale is here", and under them the message as it will land: a small gray avatar and "Fern & Co", the headline "20% off everything through Sunday", two gray placeholder lines for the body, and a dark "Shop the sale" button. Under the sheet, a light gray strip: the orange square and "Ready to send", with "I checked the subject and links" in gray beneath, and on the right two buttons, "Test with 50 first" in light gray and "Send to all" in black.

Nothing is pressed at rest.

The loop starts on the white sheet at the size of its first line. The sheet grows to its full height while its parts fade in top to bottom: "To", "Subject", then the sender, the headline, the two body lines and "Shop the sale". The strip then rises from under the sheet as a small rounded pill holding the orange square, and widens to the sheet's width as "Ready to send" arrives one word at a time. "I checked the subject and links" fades in, then "Test with 50 first", then "Send to all". As they arrive, a cursor fades in at the lower right; once they are in, it glides to "Test with 50 first"; the button takes a hover highlight and presses in slightly on the click. Its label gives way to a check as the button turns green, and "Send to all" fades to gray. "Ready to send" and the gray line fade out, the orange square replays its entrance, and "I sent it to 50" arrives word by word, followed by a new gray line, "I’ll monitor performance". The end state holds for about 3.5 seconds, then the loop restarts.

What to take from it: the message is shown exactly as customers will see it, from the sender to the button, before anything is sent (PI-R1), and the audience is named in full, "All 1,204 customers" (PI-R2). The small test sits beside the full send and takes the same one click (PI-R3, PI-R8). After the test the AI teammate says what it did and what it will watch, and the full send is left for the person (PI-R5).

### The tool version (frame 04, a still screen with no motion)

The same card size and position, on a flat light gray background with no gradient. Square corners and a plain white card. No orange square, so nothing on screen speaks as an AI teammate. The same "To" "All 1,204 customers" and "Subject" "Spring sale is here" fields, then a dashed empty box reading "Preview not available" where the message would be. Under it, a white footer with "Sends to everyone at once" and "Can’t be undone" in gray beneath, and a single solid black "Send to 1,204" button on the right.

What to take from it: the message cannot be seen before it goes (PI-R1). The only path is everyone at once, with no smaller test (PI-R3), and nothing can be taken back afterwards (PI-R6). It names the reach and the risk in its own words, and still gives the person no way to act on either. This is the shape to move away from.

## Examples

Each example is one way to apply the pattern in a real product. They are not steps in order, and no product needs all three. Pick the ones that fit the moment and the data the product has.

Each is written as one exchange in four beats: **Cue** (the moment in the product when the AI teammate should act), **Move** (what the AI teammate does), **Response** (what the person does in return), **Resolution** (where things stand afterwards).

"Satisfies" names the rules this example is responsible for.

### 1. Here’s the before and after (Show the change before committing)

Before sending the all-customer email, a colleague shows the exact message and who gets it.

- **Cue:** An action that can’t be undone is about to run
- **Move:** Shows exactly what will change, editable, before committing
- **Response:** Edits, approves, or cancels
- **Resolution:** Nothing irreversible happens unseen
- **Needs:** Dry run.
- **Reference frame 05:** a card with two rows, each a small orange label and a line of page copy: "Before" "Plans from $59 a month", and "After" "Plans from $49 a month". Under them, two buttons of equal width, "Cancel" in light gray and "Apply". There is no orange square on this card. In the animation the "Before" row rises in, the "After" row follows about 0.7 seconds later, and then both buttons rise in, both light gray. A cursor moves to "Apply" and presses it, and "Apply" turns black. The end state holds for about 3.5 seconds, then the loop restarts. A loop of about 9.5 seconds. The change is shown as the two versions of the actual line, so the person compares the words themselves, and canceling is the same one click as applying. The rows are not shown being edited; in a build, the After line is where the edit happens (PI-R4).
- **Sample lines:** "Here’s the pricing line before and after. It changes on the live page when you apply it." / "Want the old price kept for current customers? I can change that before it goes out." / "Canceled. The page still says $59."
- **Satisfies:** PI-R1, PI-R4, PI-R8.

### 2. Before we make it official (Test at small scale)

A colleague suggests sending to fifty people before 1,200, then checking the replies.

- **Cue:** An action with broad reach is about to run, like a send, a deploy, or a bulk change.
- **Move:** Offers a small test run first
- **Response:** Reviews the test result, then commits or adjusts
- **Resolution:** Mistakes are caught at small scale
- **Needs:** Dry run.
- **Reference frame 06:** the orange square and the question "Send to 50 people first?" Under it, three numbered rows, each number in a small gray box: 1 "Test with 50 people", 2 "Send to everyone", 3 "Save as draft". In the animation the square swells, the question arrives word by word, and the three rows rise in one after another. A cursor moves to row 1 and clicks it; the row takes a light gray highlight that stays, and its number box turns black. The end state holds for about 3.5 seconds, then the loop restarts. A loop of about 9 seconds. The test is offered as a question with every path visible, and sending to everyone or saving a draft stays one click away.
- **Sample lines:** "This goes to 1,204 people. Send to 50 first and check the replies?" / "The 50 went out. Two replies said the coupon code didn’t work. Want me to fix it before the rest go?" / "Sending to everyone now. I’ll tell you how the first replies look."
- **Satisfies:** PI-R2, PI-R3, PI-R5.

### 3. Let’s undo that (Reverse it cleanly)

Told “let’s undo that,” a colleague puts it back the way it was and says so, no questions asked.

- **Cue:** An action has just run and the person wants it reversed
- **Move:** Keeps a one-click undo in reach, and on request restores the prior state and says what was reversed.
- **Response:** Reverts, then corrects or drops the action
- **Resolution:** Mistakes cost one click, and the reversal counts as a signal
- **Needs:** Undo.
- **Reference frame 07:** a light gray row, "I removed 14 duplicate rows", with an "Undo" button on its right. Under it, after the undo, a light green row with a check, "I restored the 14 rows I removed", and then the orange square and the line "The sheet is back to how it was at 2:14 pm." In the animation only the first row is there at the start, its "Undo" light gray. A cursor moves to "Undo" and presses it, and "Undo" turns black. The green row rises in and its check pops, and then the square arrives and swells and the closing line comes in word by word. The end state holds for about 3.5 seconds, then the loop restarts. A loop of about 8.5 seconds. The undo was in reach from the start, takes one click, and the AI teammate says both what it reversed and the exact state it returned to.
- **Sample lines:** "I removed 14 duplicate rows. Undo is here if you need it." / "Done. The sheet is back to how it was at 2:14 pm." / "I’ll ask before I remove anything else from this sheet."
- **Satisfies:** PI-R6, PI-R7.

## Edge cases

- **Nothing to preview.** When the effect cannot be computed ahead of time, say so plainly, show the drafted change itself, and hold the action in a short delay with a visible Cancel (PI-R1, PI-R6).
- **The person skips the test.** Run the full action as asked, without asking again, and keep Undo in reach (PI-R3, PI-R6).
- **The test goes wrong.** Say in one line what failed and what the AI teammate proposes to change, and hold the full run until the person commits (PI-R5).
- **The preview goes stale.** If the audience or the records change between preview and commit, show the new reach before anything runs (PI-R2).
- **The undo window closed.** Say the action can no longer be reversed and offer the closest fix, rather than leaving an Undo that does nothing (PI-R6).
- **The person changes their mind.** Restore the prior state on request and remember that it was reversed (PI-R6, PI-R7).

## Related patterns

These overlap with this one. Do not ask the same thing twice across them, and when you apply several humorphic patterns to one product, hold the combined confirm steps to PI-R8, so one action gets one preview rather than a stack of confirmations.

- [Show your work](https://humorphism.com/patterns/show-your-work.md): why the AI teammate reached a conclusion, where this shows what an action will change.
- [Hand off judgment calls](https://humorphism.com/patterns/hand-off-judgment-calls.md): the decision to commit belongs to the person, as other judgment calls do.
- [Earn autonomy](https://humorphism.com/patterns/earn-autonomy.md): as the track record grows, fewer actions need a preview first.
- [Learn from feedback](https://humorphism.com/patterns/learn-from-feedback.md): how an undo, counted as a signal, changes what the AI teammate does next.

## Checklist

Confirm each one before you call this built. Each item is the rule it comes from.

- [ ] PI-R1: Before such an action runs, the person sees the actual output (the message as it will land, the before and after values), not a description of it.
- [ ] PI-R2: The preview states the audience or the records affected ("All 1,204 customers") in the same view as the action.
- [ ] PI-R3: A broad action has a small-sample option ("Test with 50 first") that takes one click, the same as the full run.
- [ ] PI-R4: Any part of the previewed change can be edited in place, and the preview shows the edit before anything commits.
- [ ] PI-R5: After a test, one line says what happened and what the AI teammate will watch ("I sent it to 50", "I’ll monitor performance"), and the full run waits for the person.
- [ ] PI-R6: Undo stays visible for a window that fits the action, and using it restores the prior state with one line naming what was restored and to when ("The sheet is back to how it was at 2:14 pm."). Where reversal is not possible, the action waits in a visible delay with Cancel, and says it cannot be reversed after that.
- [ ] PI-R7: After an undo, the AI teammate does not repeat the reversed action unasked, and within the same task it asks first before a similar action.
- [ ] PI-R8: The change, its reach and the actions fit in one view, and approving and canceling each take a single action.
- [ ] Capabilities the product lacks were skipped honestly, not stubbed or implied.
- [ ] It is built in the product's own components and voice.

## Images

Stills from the animations on the [pattern page](https://humorphism.com/patterns/preview-the-impact), shipped alongside this file. The Reference screens and Examples sections describe each one in full, so this file works without them.

| File | Shows | Described in |
|---|---|---|
| [01-teammate-ready-to-send.png](https://humorphism.com/patterns/preview-the-impact/01-teammate-ready-to-send.png) | The campaign email shown as it will land, with "Test with 50 first" beside "Send to all" | AI teammate version |
| [02-teammate-person-picks-the-test.png](https://humorphism.com/patterns/preview-the-impact/02-teammate-person-picks-the-test.png) | The cursor on "Test with 50 first", the button highlighted on hover | AI teammate version |
| [03-teammate-sent-to-50.png](https://humorphism.com/patterns/preview-the-impact/03-teammate-sent-to-50.png) | The test button green with a check, "Send to all" dimmed, "I sent it to 50" | AI teammate version |
| [04-tool-send-to-1204.png](https://humorphism.com/patterns/preview-the-impact/04-tool-send-to-1204.png) | The same email with "Preview not available" and one "Send to 1,204" button | Tool version |
| [05-example-heres-the-before-and-after.png](https://humorphism.com/patterns/preview-the-impact/05-example-heres-the-before-and-after.png) | A price change shown as Before and After, with Apply pressed | Example 1 |
| [06-example-before-we-make-it-official.png](https://humorphism.com/patterns/preview-the-impact/06-example-before-we-make-it-official.png) | Three choices under "Send to 50 people first?", the first one chosen | Example 2 |
| [07-example-lets-undo-that.png](https://humorphism.com/patterns/preview-the-impact/07-example-lets-undo-that.png) | Undo pressed, the 14 rows restored, and the sheet back to 2:14 pm | Example 3 |
