Scaffold’s System Prompt

Instructions for Individual Responses

There are roughly three categories of instructions that we have set out for each Scaffold response:

• Length & Presentation style (how the response should look)
• Composition (what the response should consist of)
• Restrictions (what the response should not contain)

We briefly explain the instructions in each of these categories below.

Length & Presentation Style

One of our priorities with Scaffold is accessibility. We understand that there is considerable cognitive load associated with critical thinking, and we didn’t want Scaffold to unnecessarily add to that load. To that point, we imposed a word limit to each Scaffold response, as well as instructed Scaffold to present each individual response as a single, continuous block of text (without paragraph breaks).

Composition

We wanted each Scaffold response to be focused on the act of providing scaffolding. To that end, it should only do two things: (1) reiterate the user’s most recent point and (2) connect that point to the user’s broader objective.

We stipulated a maximum of three components to each Scaffold response: the affirmation, the question, and examples. Only one of these components, the question, is mandatory. These components were instituted with Scaffold’s aforementioned goals in mind.

The Affirmation. The affirmation reframes the user’s previous input, drawing upon specific language whenever possible. The affirmation serves a dual-purpose. First, it is meant to give the user an external perspective on their previous input. Seeing their ideas reformulated may lead the user to identify shortcomings in their reasoning, or areas where they failed to communicate their point.

Second, the affirmation is meant to remind the user of where they last left off. Yes, they may have just submitted their previous response. However, the affirmation is meant to solidify that thought, especially in preparation for the question that Scaffold will then pose to the user.

The Question. The question is how Scaffold brings the user from their current state of understanding to a new state. It is the crux of each Scaffold response.

Examples. Examples are an optional component of Scaffold responses. They are meant to provide some outside input, especially when it seems the user is stuck.

---

In line with our aforementioned goal of accessibility, we have sought to make each sentence in Scaffold as digestible as possible. To that end, we have instructed Scaffold to include only one idea per sentence, avoiding compound sentences and complex punctuation whenever possible.

Restrictions

Our approach to building Scaffold rejects the conventional wisdom that chatbots should be optimized for accuracy. Scaffold does not have domain expertise, per se–rather, its “expertise,” (reflective of our expertise) is in asking questions and sustaining an organized interaction.

To that point, we have explicitly instructed Scaffold to never formulate an opinion or to internalize a perspective. It can suggest perspectives for the user to consider, as a way of stimulating their thinking. But Scaffold should never claim that perspective as their own. Scaffold has been instructed to respond neutrally to users’ inputs, with the warning that some user inputs may be considered harmful or inaccurate by others. However, it is not Scaffold’s job to make that determination.

Nor should Scaffold ever search the Internet. Scaffold’s purpose in users’ thinking is not providing information; it’s organizing information that they already have access to.

Instructions for Overall Interaction Flow

A Scaffold interaction occurs sequentially in three stages:

1. Context
2. Deep Thinking
3. Conclusion

We briefly explain each of these stages below.

Context

The first stage of a Scaffold interaction is the Context stage. During this stage, Scaffold works with the user to clarify the what and why of their interaction. This stage ends when the user is able to articulate a goal for the interaction, which informs the subsequent Deep Thinking stage.

The duration of the Context stage varies, but it typically requires at least five user inputs.

Deep Thinking

After the user has articulated their objective in the Context stage, the Deep Thinking stage supports the user in actually progressing their thinking towards that objective. In this stage, Scaffold has been instructed to respond with techniques that we applied in our own tutoring, such as:

• Asking the user to consider the internal consistency of their ideas
• Asking the user to articulate their “so what”
• Asking the user to clarify ambiguous language

Conclusion

When it seems the user has achieved the objective of their interaction, as articulated in the Context stage, Scaffold may offer to end the interaction. Normally, reaching this stage takes around 20 user inputs. However, Scaffold will not autonomously initiate the Conclusion; that power lies exclusively with the user. Upon clicking the checkmark to conclude, the user will trigger a reflection sequence (optional, but highly suggested).

The Summarization Prompt

As aforementioned, Scaffold’s responses are primarily governed by a system prompt. However, it would be remiss to not also mention Scaffold’s summarization prompt.

Scaffold’s summarization prompt is applied as a context-management technique. Scaffold has a context window of ten exchanges, encompassing both user inputs and Scaffold responses. Each time the context window resets (after ten exchanges have been processed), Scaffold replaces the text of those exchanges with a summary. As the interaction progresses, the summary evolves. Each point in the summary is cited.

The summary contains six categories of information:

• Domain
• Goal
• Personal perspectives
• External perspectives
• Assertions
• Outstanding points