---
title: "Ask before acting"
pattern_id: "ask-before-acting"
trait: "Invested"
type: "Humorphic pattern brief for coding agents"
spec_version: 3
canonical_url: "https://humorphism.com/patterns/ask-before-acting"
markdown_url: "https://humorphism.com/patterns/ask-before-acting.md"
index: "https://humorphism.com/llms.txt"
all_patterns_in_one_file: "https://humorphism.com/llms-full.txt"
normative_section: "Rules"
rule_prefix: "AA"
rule_ids: ["AA-R1", "AA-R2", "AA-R3", "AA-R4", "AA-R5", "AA-R6", "AA-R7", "AA-R8", "AA-R9"]
---

# Ask before acting

## 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 ask the one or two questions that would change the outcome before it starts anything consequential, and ask nothing else. 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: the moment a request arrives that would start costly or hard-to-undo work (a build, a publish, a send, a bulk change), a long or ambiguous brief, and any task where a wrong start costs more than a short pause.
2. **Read Rules.** All nine 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.** If you build only one, build Example 1.
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: Invested. Aligns with the person’s goals and cares whether they succeed_

Colleagues don’t build on a guess. A good AI teammate asks before anything consequential, and asks only what would change the work.

**Build this.** An AI teammate asks the one or two questions that would change the outcome, because a wrong start costs more than a short pause.

**Avoid this.** A tool starts building on the first prompt, because it treats every request as complete.

The difference to carry into your build:

| Tool | AI teammate |
|---|---|
| Treats the first prompt as complete | Spots the one or two unknowns that would change the work |
| Makes every open decision silently | Asks about the decision that changes the outcome most |
| Asks nothing, or asks everything | Asks only what would change the work, and says why |
| Assumptions are hidden in the result | Assumptions are stated up front and can be changed mid-flight |
| A wrong start is found after the work is done | A misreading is fixed before any work happens |

## Rules

Normative. All nine 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 |
|---|---|---|---|
| **AA-R1** | Ask only before consequential work. Start small, easily undone requests without asking. | A question on every request turns a short pause into friction the person learns to skip. | Small, reversible requests start at once, and questions appear only where a wrong start is costly to undo. |
| **AA-R2** | Ask only the questions whose answers would change the work. | A question whose answer changes nothing costs the person time and teaches them the questions do not matter. | Every question asked, if answered differently, would produce different work. |
| **AA-R3** | Say why each question matters, in one short line. | A reason shows the question is about the work, not a form to fill. | Each question carries a one-line reason naming what it changes ("Checkout or phone orders changes the whole build."). |
| **AA-R4** | Make each question answerable in one action, with the likely answers offered and room for the person's own. | Offered answers make the pause seconds long instead of a writing task. | The person can answer with one click on an offered answer, or type a different one. |
| **AA-R5** | Restate a long, ambiguous, or high-stakes ask in one or two lines before starting, with each part correctable in place. | A misreading caught in one line costs nothing; caught in the finished work it costs the work. | Before big work begins, one or two lines name the goal, scope, and constraints as heard, and the person can fix any part of that line directly. |
| **AA-R6** | Let the person skip the questions, or leave them unanswered, and name the assumptions you will proceed on before you start. | A required answer blocks the work; a silent guess hides it. | A "go ahead" path is always open, and when it is taken the AI teammate lists the assumptions it is running on before the work starts. When a question goes unanswered, nothing that cannot be undone, such as a publish or a send, runs until the person answers or says go ahead. |
| **AA-R7** | Let the answer change the work. What starts after the answer must follow from it. | A question whose answer is ignored is the tool version with an extra step. | The work that begins after an answer names that answer and differs by it ("Got it. Building online checkout…"). |
| **AA-R8** | Keep every assumption visible and changeable while the work runs. | An assumption the person cannot reach becomes a wrong result they find too late. | The assumptions stay on screen during the work, and changing one updates the work from that point without restarting it. |
| **AA-R9** | Keep it short. One or two questions in a single turn, then start. | A run of questions is a form in a chat's clothes, and it delays the work it was meant to protect. | No request gets more than two questions before work begins, and they arrive together in one turn. |

**AA-R7 carries the pattern.** Wire each answer into the work it changes before you polish how the question is asked. A build that asks well-chosen questions and then does the same thing whatever the person says has produced a delay with good manners: it looks like the pattern and behaves like the tool version.

## 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 |
|---|---|---|---|
| Autonomy policy | Settings for how the AI teammate acts in different scenarios: ask first, do then tell, or do silently. The person can change them, and every past approval and reversal is on record. | Example 3 | Ask before every consequential request and always offer "go ahead on your best guess" as one of the answers. Do not claim the person has set a standing choice about when to ask. |
| Memory | A place that keeps facts, preferences, rules, and open threads across sessions. The person can see everything the AI teammate remembers and edit or delete any of it. | Example 3 | Hold the stated assumptions for the current task only, and say so. Do not claim to remember them, or the person's answers, next time. |

Examples 1 and 2 need none of these. **If you build only one example, build Example 1.**

## 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 the person's message sits in a gray bubble on the right, the AI teammate's lines sit on the left with no bubble, and a small orange square marks the AI teammate's side: under its latest turn in the chat, and at the head of its line in the example cards. The AI teammate's side sits on a warm, slowly moving gradient while the tool's side sits on flat gray. Offered answers are outlined chips or light gray buttons; the person's choice registers as a change of state (the chosen chip fills gray and the others fade, or the chosen button turns black), and a correction is made in place, with the old value struck through and the new one typed beside it. Use whatever your system already uses for "the AI teammate is speaking", "these are suggested answers", "this is chosen" and "this was edited".

### The AI teammate version (frames 01 to 03, a loop of about 10 seconds)

One chat window, centered on a warm, slowly moving gradient of pink, orange and blue. At the top right, the person's message in a light gray bubble: "Make me a site for my bakery". Below it on the left, the AI teammate's reply as a single line of text with no bubble: "Before I build: how do people order?" Under the line, three outlined chips in a row: "Online checkout", "Call or visit" and "Both". The orange square sits under the chips, marking the AI teammate's turn. At the bottom, an empty composer with the placeholder "Message" and a round black send button.

Nothing is selected at rest, and nothing is being built.

The loop starts on an empty chat window. The person's bubble, "Make me a site for my bakery", fades in at the top right. A beat later the orange square appears on the left and the question arrives one word at a time just above it; as the words finish, the square slides down and the three chips fade in one after another, left to right, so the square comes to rest under them. The cursor waits at the lower right, then glides up to "Online checkout"; the chip fills light gray on hover and presses in slightly on the click. "Online checkout" stays filled, "Call or visit" and "Both" fade to a lighter gray, and the orange square moves down to start a new turn, where "Got it. Building online checkout…" arrives word by word; then the cursor fades away. The end state holds for about 3.5 seconds, then the loop restarts on the empty chat window.

What to take from it: the AI teammate asks one question, the one that decides the whole build, and asks it before anything is built (AA-R1, AA-R2, AA-R9). The answer is one press on an offered chip (AA-R4). And the reply that starts the work names the answer it was given, so the build follows from the choice (AA-R7). Two things this screen leaves out still apply: the one-line reason under the question (AA-R3, shown in Example 1) and a way to skip the question and go ahead on stated assumptions (AA-R6, shown in Example 3).

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

The same chat window, square-cornered, on a flat light gray background with no gradient. No orange square, so nothing on screen speaks as an AI teammate. The person's bubble reads "Make me a site for my bakery". Under it, the reply "Building your site…", then a two-column checklist of build steps separated by thin dividers, each with a small box. Six are ticked: "Online checkout", "Stripe payments", "Orange palette", "Serif headings", "14 pages written" and "Blog section". Two are still running, with empty dashed boxes and lighter text: "Newsletter pop-up…" and "Publishing…". The composer reads "Message".

There is no question, no stated assumption, and no way to change any of the decisions on screen.

What to take from it: the tool treats a one-line request as complete and starts consequential work on it (AA-R1, AA-R2), makes every open decision, from payments to the palette, without saying it is assuming anything (AA-R6), and leaves the person nothing to correct while it runs (AA-R8). It is already publishing. 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. Quick question before I start (Ask only what changes the outcome)

A colleague handed a brief asks the one or two things that would change the direction, and nothing more.

- **Cue:** A request arrives with one or two unknowns that would send the work in a different direction.
- **Move:** Asks only the questions that would change the outcome, and says why each matters
- **Response:** Answers, or tells the AI to go ahead on its best guess
- **Resolution:** Work begins pointed in the right direction
- **Needs:** nothing beyond the AI teammate itself.
- **Reference frame 05:** the orange square and the line "Before I build anything: how do people order?" at the top of a white card on flat gray. Under it, in smaller gray text, the reason: "Checkout or phone orders changes the whole build." Under that, two buttons of equal width side by side, "Phone orders" and "Online checkout". In the animation the empty card appears with the square, the question arrives word by word, the reason rises in beneath it, and the two buttons rise in, both light gray. The cursor moves up from the lower right to "Online checkout" and presses it, and the button turns black. The card holds on the chosen answer for about 3.5 seconds, then the loop restarts. One question, its reason in one line, and two answers that each lead to a different build.
- **Sample lines:** "Before I build anything: how do people order? Checkout or phone orders changes the whole build." / "Got it. I’ll build online checkout and leave phone orders off the site." / "No answer yet, so I’ll start with online checkout, since it covers both pickup and delivery, and you can switch it anytime."
- **Satisfies:** AA-R1, AA-R2, AA-R3, AA-R4, AA-R6, AA-R7, AA-R9.

### 2. Just to be clear (Restate the ask)

Before big work, a colleague says the brief back in their own words so both sides hear the same target.

- **Cue:** The request is long, ambiguous, or high-stakes.
- **Move:** Restates the goal, scope, and constraints in one or two lines
- **Response:** Confirms or fixes the misreading before any work happens
- **Resolution:** Both parties share the same target; wasted work is avoided
- **Needs:** nothing beyond the AI teammate itself.
- **Reference frame 06:** the orange square and the line "Got it, live by June 14." at the top of a white card. Under it, the ask read back as one sentence: "A 5-page site with online orders, live by May 30 June 14." The parts that define the work sit on light gray tokens: "5-page site", "online orders", and the date, where "May 30" is struck through, "June 14" follows it with a text cursor after it, and the token has a dark outline. In the animation the square appears and "Here’s what I heard." arrives word by word as the header. The sentence rises in beneath it, and its three tokens pop in one after another, the date reading "May 30". The cursor moves to the date and clicks it; the token takes the dark outline, "May 30" is struck through, and "June 14" appears after it with a cursor. "Here’s what I heard." then fades out, the square pulses, and "Got it, live by June 14." arrives word by word in its place. The card holds for about 3.5 seconds, then the loop restarts. The misreading is caught and fixed inside the restatement, before any work starts.
- **Sample lines:** "Just to be clear: a 5-page site with online orders, live by May 30. Did I get that right?" / "Got it, live by June 14. I’ll plan the pages around that date." / "You haven’t confirmed yet, so I’ll start on the parts that don’t depend on the date and hold the rest."
- **Satisfies:** AA-R1, AA-R5, AA-R9.

### 3. Here’s what I’ll assume (State the defaults)

When there’s no time for questions, a colleague names the assumptions they’re running on and gets going.

- **Cue:** The person defers or skips the questions
- **Move:** Names the assumptions it will proceed on, then starts
- **Response:** Lets it run, or corrects an assumption mid-flight
- **Resolution:** Momentum is preserved and every assumption is visible and reversible
- **Needs:** Autonomy policy and Memory.
- **Reference frame 07:** the orange square and the line "Here’s what I’ll assume, unless you say otherwise." at the top of a white card. Under it, three light gray chips in a row: "USD pricing", "Mobile first" and "Dark theme". Under the chips, a light gray row reading "Building the site" with an orange progress bar beside it. In the animation the line arrives word by word, the three chips pop in one after another with the third reading "Light theme", and the "Building the site" row rises in beneath them, its orange bar already full. The cursor moves to "Light theme" and clicks it, and the chip's label changes to "Dark theme" while the build row stays where it is. The card holds for about 3.5 seconds, then the loop restarts. The work starts on assumptions the person can see, and one is changed mid-flight without stopping the work.
- **Sample lines:** "Here’s what I’ll assume, unless you say otherwise: USD pricing, mobile first, light theme." / "Switched to a dark theme. Everything else carries on as it was." / "You didn’t change anything, so I’m building on these. You can change any of them while I work."
- **Satisfies:** AA-R6, AA-R7, AA-R8.

## Edge cases

- **Nothing worth asking.** When the request already settles everything that would change the work, start without a question (AA-R2).
- **A vague request.** When the request names almost nothing, restate the most likely reading in one line and ask only for the part that would change it most (AA-R5, AA-R9).
- **The person skips.** When the person says "just go", start on the assumptions named out loud, never on a silent guess. When they do not answer, name the assumptions and start only the parts that can be changed later, as Examples 1 and 2 do (AA-R6).
- **A wrong assumption.** When an assumption turns out wrong while the work runs, let the person change it in place and carry the change forward from that point (AA-R8).
- **A changed answer.** When the person changes an earlier answer after work has started, treat it as a new answer: the work from that point follows it, without starting over (AA-R7, AA-R8).

## 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, count every question the AI teammate asks before work starts across all of them, and hold that total to AA-R9 rather than each pattern on its own.

- [Understand the person’s goals](https://humorphism.com/patterns/understand-the-persons-goals.md): learns the goal once, so the questions asked here can be only about this piece of work.
- [Preview the impact](https://humorphism.com/patterns/preview-the-impact.md): shows the finished result before it is applied, a check once the work is done, where this pattern checks before it starts.
- [Earn autonomy](https://humorphism.com/patterns/earn-autonomy.md): decides when the AI teammate can stop asking and act, as trust builds.
- [Push back when it matters](https://humorphism.com/patterns/push-back-when-it-matters.md): speaks up when the answer itself looks risky, rather than asking what it is.

## Checklist

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

- [ ] AA-R1: Small, reversible requests start at once, and questions appear only where a wrong start is costly to undo.
- [ ] AA-R2: Every question asked, if answered differently, would produce different work.
- [ ] AA-R3: Each question carries a one-line reason naming what it changes ("Checkout or phone orders changes the whole build.").
- [ ] AA-R4: The person can answer with one click on an offered answer, or type a different one.
- [ ] AA-R5: Before big work begins, one or two lines name the goal, scope, and constraints as heard, and the person can fix any part of that line directly.
- [ ] AA-R6: A "go ahead" path is always open, and when it is taken the AI teammate lists the assumptions it is running on before the work starts. When a question goes unanswered, nothing that cannot be undone, such as a publish or a send, runs until the person answers or says go ahead.
- [ ] AA-R7: The work that begins after an answer names that answer and differs by it ("Got it. Building online checkout…").
- [ ] AA-R8: The assumptions stay on screen during the work, and changing one updates the work from that point without restarting it.
- [ ] AA-R9: No request gets more than two questions before work begins, and they arrive together in one turn.
- [ ] 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/ask-before-acting), 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-asks.png](https://humorphism.com/patterns/ask-before-acting/01-teammate-asks.png) | The one question, "Before I build: how do people order?", with three answers to press | AI teammate version |
| [02-teammate-person-answers.png](https://humorphism.com/patterns/ask-before-acting/02-teammate-person-answers.png) | The cursor pressing "Online checkout", the chip gray on hover | AI teammate version |
| [03-teammate-builds-what-was-chosen.png](https://humorphism.com/patterns/ask-before-acting/03-teammate-builds-what-was-chosen.png) | "Online checkout" chosen, the others faded, and "Got it. Building online checkout…" | AI teammate version |
| [04-tool-builds-on-a-guess.png](https://humorphism.com/patterns/ask-before-acting/04-tool-builds-on-a-guess.png) | The same request answered with a build already eight decisions in | Tool version |
| [05-example-quick-question-before-i-start.png](https://humorphism.com/patterns/ask-before-acting/05-example-quick-question-before-i-start.png) | The ordering question with its reason, and "Online checkout" chosen | Example 1 |
| [06-example-just-to-be-clear.png](https://humorphism.com/patterns/ask-before-acting/06-example-just-to-be-clear.png) | The ask read back as one line, the date fixed from "May 30" to "June 14" in place | Example 2 |
| [07-example-heres-what-ill-assume.png](https://humorphism.com/patterns/ask-before-acting/07-example-heres-what-ill-assume.png) | Three stated assumptions, "Dark theme" just changed, with the build still running | Example 3 |
