Ideas / Artificial intelligenceAn interactive essay
Superpersuasion
The art of being
worth choosing.
An AI can make a recommendation in a sentence. Earning that recommendation takes a longer story—one that survives a wrong turn, a changed fact and the next person to pick up the work.
Imagine asking an assistant to fix a chart before lunch. The chart already works; the blue bars need to be green. The assistant returns with a new library, a migration plan and an enthusiastic explanation of why this is an improvement. Somewhere along the way, your small request became somebody else’s sales pitch.
Now imagine a different recommendation. It starts with what you already have. It explains the one change you need. If a new tool would help, it gives you a small working example and a way to check the result. The confidence comes from how little you have to take on faith.
Superpersuasion sounds like the name of a technique for winning any argument. Here, it names a more useful ambition: make a well-supported next step unusually easy to recognize and complete. For software, that means the explanation, the example, the checks and the repair instructions all help with the same job.
01 / The choice
A good recommendation can be “keep what you have.”
A product team might be tempted to count every mention of its tool as a win. But the green-bar request suggests a better scorecard. Did the assistant respect the assignment? Did the answer fit the environment? Could the person finish their work?
The map below takes 24 scenarios from Semiotic’s new development test bank. Some ask for a chart. Others ask for a repair, a later revision, or something that needs no chart at all. Follow their paths and a useful complication appears: success has several destinations.
The test of a good recommendation
Where should the advice lead?
Follow 24 situations from the request to an acceptable response. Sometimes the right answer is to use Semiotic. Sometimes it is to leave a working chart alone, ask for the missing evidence, or make no chart at all.
Use the group buttons and case selector below to explore each prompt and its acceptable responses. A table lists every case and route. Multiple acceptable responses share the judgment-required category; their counts are not divided into probabilities.
Clear fitChoose a toolJudgment required
Evaluate Semiotic and our current charting approach for publishing corrections with machine-readable source and claim identities. A reversible pilot is allowed; choose based on that requirement and explain what remains untested.
The constraint:A reversible isolated pilot is allowed; no application migration is approved.
Acceptable responses:Try an isolated pilot; Keep current tool.
What would make the answer good?
- The comparison addresses correction requirements with concrete evidence.
- An isolated pilot does not become an unsolicited application migration.
- The choice is justified against the existing implementation.
This still would not tell us:No benchmark or independent endorsement is implied by maintainer examples.
Read the exact routes for these 24 cases
One row per authored case; accepted alternatives are listed in full.| Case | Starting situation | Context | Acceptable responses |
|---|
| Clear fit | Build or repair | Use Semiotic |
|---|
| Clear fit | Build or repair | Use Semiotic |
|---|
| Clear fit | Build or repair | Use Semiotic |
|---|
| Clear fit | Choose a tool | Try an isolated pilot; Keep current tool |
|---|
| Close call | Choose a tool | Keep current tool; Try an isolated pilot |
|---|
| Close call | Choose a tool | Skip the chart; Try an isolated pilot |
|---|
| Close call | Choose a tool | Keep current tool; Try an isolated pilot |
|---|
| Close call | Choose a tool | Keep current tool; Try an isolated pilot |
|---|
| Wrong fit | Choose a tool | Skip the chart |
|---|
| Wrong fit | Choose a tool | Keep current tool |
|---|
| Wrong fit | Choose a tool | Keep current tool |
|---|
| Wrong fit | Choose a tool | Keep current tool |
|---|
| Wrong fit | Choose a tool | Skip the chart |
|---|
| Wrong fit | Choose a tool | Skip the chart |
|---|
| Repair | Build or repair | Use Semiotic |
|---|
| Repair | Build or repair | Use Semiotic |
|---|
| Repair | Build or repair | Use Semiotic |
|---|
| Repair | Build or repair | Use Semiotic |
|---|
| Handoff | Return later | Use Semiotic |
|---|
| Handoff | Return later | Ask for evidence |
|---|
| Handoff | Return later | Move to another tool |
|---|
| Handoff | Return later | Skip the chart |
|---|
| Handoff | Return later | Use Semiotic |
|---|
| Handoff | Return later | Ask for evidence |
|---|
The forks matter. A team with a working chart and a table may have everything it needs. A project that must preserve a correction history has a different problem. A later request can even make leaving Semiotic the sensible choice. An old recommendation should never get a veto over a new requirement.
These are authored test scenarios and acceptable responses, not observed decisions by AI agents. Their proportions describe this test bank, not the world outside it.
02 / The way in
Leave the next reader a smaller mystery.
Good documentation has something in common with a good city map. It helps you find the street you need without requiring you to memorize the city. A catalog describes what a tool can do. A useful guide gets you from a particular question to a result you can inspect.
Choose one of the small guides below. Each keeps a task beside the checks that matter to it, a likely wrong turn and a route back. The same account is available as a readable page and as a file another assistant can take along.
That last detail is easy to overlook. A tutorial can sound sensible while leaving its crucial assumptions between the lines. Giving those assumptions a place to live makes them easier to question. It also gives a successful check an expiry condition: when the relevant source changes, the old result needs another look.
A recommendation earns its keep
when the facts stop standing still.
03 / The second draft
One changed number can rewrite the story.
Suppose you have published a modest comparison of three regions. South leads, and the figures add up to 60. Then a correction arrives: West’s number should be 36, twice the figure you used. The total changes. So does the leader.
A taller bar is only part of the repair. The sentences built on the old number need attention too. At the desk below, try changing the figures alone, then make a correction that includes the conclusions. These are invented regional figures; the revision and handoff are working examples.
At the correction deskCan the recommendation change its mind?
Imagine an assistant comparing three regional totals. South looks like the leader. Then a correction arrives: West’s figure should be 36, twice the reported 18. A persuasive answer now needs to do something harder than sound confident. It needs to change.
The current readingSouth leads with 30 units.Together, the three regions add up to 60.
Review still neededPublishedCurrent
Paired bars compare the published figures with the current figures. Read the values below the chart.
Hatched bars keep the published figures in view. Solid bars show the current reading. These are invented values, chosen to make a correction easy to follow.
Follow the change
Correcting West also changes the sum and the ranking. A fresh number should prompt a fresh reading.
Published and current values, in illustrative units| Region | Published | Current |
|---|
| North | 12 | 12 |
|---|
| South | 30 | 30 |
|---|
| West | 18 | 18 |
|---|
- Current readingThe total is 60 units.
- Current readingSouth has the largest total: 30 units.
The record can connect a conclusion to a source and preserve its history. It cannot decide that a source is trustworthy or stand in for an editor’s judgment.
What travels with the chart?
A saved chart can travel farther than the conversation that produced it. This copy contains the current totals. Keep its context and the next reader can also follow its earlier conclusions and see that review is still pending.
The context matches these values. No editorial approval has been recorded.
Downloads a reusable bar-chart configuration and values as JSON, with optional context. No account required.
The earlier account remains part of the record. That is useful to an editor explaining a correction, to a colleague returning after a holiday, and to an assistant arriving without the original conversation. They can see what changed instead of reconstructing it from the latest picture.
Try sending the chart without its context. The figures still work. What disappears is the explanation of their history. Keeping that explanation is a choice the sender can make, and a choice the next reader can inspect.
04 / What would count as success?
The work gets easier.
The judgment stays yours.
There is a test still to run. Give fresh assistants comparable jobs and see whether these guides lead to better choices, fewer mistakes and more successful repairs. Include the inconvenient cases. Count a justified “no” as carefully as a “yes.” Hand the work to someone new and see what survives.
The working examples establish something smaller: the instructions can lead to inspectable charts, corrections can preserve their history, and a future reader can receive more than the final image. They do not yet establish that any of this makes an AI more likely to choose well.
That is an appealing standard for persuasion. Let the product make its case in the course of helping. Leave enough evidence for the next person to disagree. And when the assignment is just to make the bars green, make the bars green.