From Prompting to Conversation Design: Managing Context in Long AI Conversations

 Conversation Modes as Intent Framing

I use Generative AI tools regularly for a variety of tasks, to the point where I’ve established a simple taxonomy of conversation modes. The modes are: 1) Teaching; 2) Research/Validation; 3) Writing/Creative; and 4) Review/Audit. Each conversation mode represents a distinct conversation goal.

In Teaching Mode, the goal is to explain, to build some scaffolding, to clarify a concept, or to identify relevant examples.

In Research/Validation Mode, the goal is to check claims, separate evidence from inference, and surface uncertainty.

In Writing/Creative Mode, the goal is to develop ideas, framing, outlines, work with voice, and explore creative possibilities.

In Review/Audit Mode, the goal is to diagnose strengths, gaps, risks, and alternatives.

The purpose of defining these modes is to more clearly guide the AI by sharing my conversation intent upfront. I’ve embedded these conversation modes in ChatGPT Plus custom instructions. The result is that if I start a creative writing conversation, the AI focuses on the goals of creative writing and does not bring in unrelated context from other types of work.

Why Long AI Conversations Break Down

Regardless of the mode selected, and acknowledging that modes can overlap, longer AI conversations tend to degrade over time. The issue is unmanaged context. Selecting a conversation mode narrows context at the outset, but many of the most valuable conversations are meandering. They begin with one intent and evolve unpredictably.

Most current guidance focuses on prompting, which assumes short exchanges and treats iteration as a sequence of follow-up prompts to improve responses. Prompting alone does not address the dynamics of extended conversations.

How can we design conversations with AI?

Three Failure Modes

I’ve experienced three distinct failure modes in long conversations.

Path loss: Each turn commits to one direction among several possible paths. Alternative directions are left behind and become difficult to recover. At times I pursue a clarification knowing it is a temporary detour, while the AI interprets it as a new primary direction.

Context overload: Over multiple turns, accumulated content becomes unstructured and difficult to navigate.

Premature convergence: Following a single linear path, often reinforced by the AI’s tendency to affirm coherent reasoning, can lead to closing a conversation without exploring alternatives or validating key assumptions.

Context as the Missing Structure

If context is the underlying issue, the question becomes how to make it visible and manageable.

The paper Mixed-Initiative Context: Structuring and Managing Context for Human-AI Collaboration explores this problem in depth. It treats conversation context as something that can be actively structured and manipulated rather than passively accumulated. The authors propose an interface where conversations can branch into subtopics, where parts of the history can be selectively excluded, and where both the user and the AI participate in organizing the interaction.

The approach resonated with me because of its close relationship to concept mapping, insight mapping, and conversation mapping, which I continue to explore. Without access to such an interface, I began to consider how a similar approach could be approximated within a standard chat environment.

Approximating Structure in Standard Chat

This led me to develop a lightweight conversation protocol. I do not rely on fixed commands or special syntax. These are natural language interventions used selectively when the conversation begins to drift.

Conversational parking lot: I might say, “Park: Draft a blog post about this topic.” This captures and defers a thread without losing it.

Validation hold: I might say, “Let’s return to this argument for validation.” This marks a claim or direction that requires verification before proceeding.

Summarization checkpoints: I often summarize at the end of a conversation, especially to capture insights in my Personal Knowledge Management system. My PKM uses TiddlyWiki, and I have set instructions so that when I enter (tw:), the output is formatted appropriately. Summaries can also be used mid-conversation. For example: “Summarize the conversation so far with a focus on what we’ve decided to pursue,” or “List any parked items.”

I do not expect the AI to decide what to park, validate, or summarize. I remain responsible for guiding the conversation. The AI assists by structuring and tracking what has been identified.

A Short Example in Practice

In earlier reflections on my use of GenAI, I noted the need to slow down and introduce pauses. This protocol adds more specific types of pauses.

Here is a shortened example, showing the intervention points:

User: I’m thinking about a blog post on improving long AI conversations. It might connect to that paper on context.

AI: You could frame it around prompting techniques, or focus on conversation structure.

User: Yes, structure feels right. Though I also want to include a concrete example, maybe from my own use. (Conversation continues for a few turns, mixing ideas)

User: Pause and review. I think we have two threads: 1. framing around conversation design 2. example from practice Let’s focus on the first. Park the example for later.

AI:
• Active: framing conversation design
• Parked: example from practice

User: I want to bring in the paper, but I am not sure I am representing it accurately. Let’s put a validation hold on the paper’s main claim before we build around it.

AI: Here is a concise restatement of the paper’s core idea. (Conversation continues, now more focused)

User: Pause and summarize what we’ve established so far.

AI:
• Framing: shift from prompting to conversation design
• Mechanisms: parking, validation, summarization
• Pending: example (parked), confirm interpretation of paper

The conversation has not become simpler. It has become more structured.

Limits and Tradeoffs

I will need to test this approach over time. I tend to introduce structure, but not all of it persists. If the cognitive overhead is too high or the benefits are not clear, I may not apply the protocol consistently.

This approach requires discipline and intentionality. It is not necessary for short or low-stakes interactions. If the need does not arise frequently, the practice may fade. This is a minimal attempt to address some of the challenges discussed in the paper within the constraints of current tools.

From Prompting to Conversation Design

I continue to teach prompting, and it remains a necessary skill. For extended interactions, prompting alone is insufficient. As conversations grow in length and complexity, users must actively manage context. This reflects a shift from crafting individual inputs to shaping the trajectory of a conversation.

Given the pace of change in AI tools, this approach may have a limited lifespan. Interfaces may evolve to incorporate these capabilities directly. For now, it offers a practical way to improve the quality of long, exploratory conversations.

Experiment: Mapping Cognitive Agency Under AI Amplification

Over the past few days, I conducted a small experiment in hybrid concept mapping. The goal was straightforward: take a loose cluster of existing notes from my TiddlyWiki and transform them into a structured map that could withstand scrutiny.  These notes (42 of them) were all tagged "LS-Cognitive". The LS stands for Learning Sprint.  

The working title of the resulting draft map is:

Maintaining Cognitive Agency and Brain Capital Under Conditions of AI Amplification

The process unfolded in stages.

First, I extracted key concepts from existing tiddlers and generated a provisional graph structure using AI (Ace Knowledge Graph, accessible as an app within ChatGPT Plus). This initial pass surfaced nodes such as Cognitive Agency, Cognitive Control, Cognitive Offloading, Human–Machine Symbiosis, Brain Capital, and Brain Economy. The resulting structure was technically coherent but conceptually uneven. This is the first time I manage to do something like this combining my TiddlyWiki content with mapping so I was impressed no matter how imperfect the map was. 

The Ace Knowledge Graph performed three useful operations:

1. Entity Extraction
   It identified recurring conceptual terms across the tiddlers and normalized them into discrete nodes.

2. Relationship Inference
   It generated directional edges based on linguistic and semantic cues in the text. These edges were labeled with descriptive phrases derived from the underlying content.

3. Structural Flattening
   It produced a graph representation that made implicit assumptions visible. Concepts that felt intuitively connected were now linked explicitly.

What Ace did not do was evaluate coherence. It did not ask whether the abstraction levels were aligned, whether certain nodes were redundant, or whether a central thesis existed. It generated a plausible structure.

Next, I imported the structure into CmapTools (my favorite concept mapping software) using a simple tab-delimited text file (other import formats failed). Here again, AI did then necessary format conversions.  The import preserved linking phrases and allowed manual rearrangement. Eliminating crossing lines forced clarification of hierarchy and exposed abstraction mismatches.  The node/concept definitions were manually added as info notes. 

At that point, the real work began. I removed some nodes, questioned ambiguous terms such as “learnability,” and examined whether each concept genuinely contributed to the central thesis. Some terms migrated to side modules. “Aging Well,” for example, remains connected by a dotted line, signaling a future expansion into lifespan cognition rather than a core dependency.  Not surprisingly, the map isn't entirely coherent because the original notes were collected over time without any concern around a central question or well-defined Learning Sprint. They only had one thing in common, they were related to "cognition". 

This map is not intended as a comprehensive theory. It functions as an interrogation device. Each node can be a starting point for further reading and research as well as future elaborations of the map. The map evolves as a record of structured reflection rather than a finished system.

Going forward, this experiment could become a template. Each Learning Sprint may generate its own evolving concept map, versioned and periodically reassessed. The value lies in making the architecture of thinking visible and open to revision.

The Contribution Spectrum: From General Volunteer to Pro Bono Work

Most of what I write publicly is posted in one or the other of my Substacks.  However, once in a while, a topic doesn't belong in either of those and I post it here.

Martin Luther King Day of Service
This piece grew out of an in-person conversation on January 16th, 2026 at a meeting of the Knowledge Management Community of DC (KMCDC) held around the Martin Luther King Jr. Day of Service at the MLK Library in Washington, D.C. I am grateful for Ninez Piezas-Jerbi for organizing and for the contributions of all the participants. The following also includes post-session reflections and as such, it is not a summary of the discussions per se. 

The session focused on volunteering and pro bono work, and on the many ways people contribute time, skills, and expertise outside of paid roles.

What became clear in the discussion is that the usual distinction between “volunteering” and “pro bono” is too blunt to be very helpful, especially for knowledge workers. Most opportunities fall somewhere in between, and the differences that matter are not only about skill level, but also about structure, motivation, accountability, and impact.

The notes that follow reflect an initial attempt to make those distinctions more explicit. They are not a prescription, but a framework for thinking more carefully about fit, expectations, and tradeoffs when deciding how to contribute. 

Infographic generated by AI (NotebookLM) based on the text of this post, which is human-generated.

Volunteer vs. pro bono: What's the difference?

Type of Task

There's a spectrum of opportunities for doing “work” that does not fall into the category of a paid position, contract work, internships, or fellowships.

At one extreme, a ''general volunteering'' opportunity may involve serving food at a soup kitchen or participating in a community stream clean-up. Unless you're a chef designing the menu being served at the soup kitchen, your skill set is not especially relevant to the task, but you are providing valuable labor. This is not to say that general volunteering does not require skills. Working with other volunteers, and being a volunteer, often requires social skills and restraint, regardless of the task.

At the other extreme of this spectrum, you may find a lawyer providing ''pro bono services'' to a client. In such cases, the expertise of the person providing pro bono services is central to the work. Pro bono tax advising would fall close on this end of the spectrum as well.

In between these two extremes, there are many opportunities to contribute that sometimes mix the two modes, including ''skill-based volunteering.''

General volunteering tasks may be very scripted and specific, especially when undertaken with organizations that manage large numbers of volunteers and know how best to leverage their time and energy. Tasks are clearly assigned. You accomplish the task as defined. There may be little room to apply your expertise, even if you believe it is relevant. KM practitioners, in particular, may see every small dysfunction as a KM problem that could be addressed with a KM solution. Apply restraint here unless you have been specifically asked for feedback.

What's your Motivation?

Why are you volunteering or doing pro bono work? Motivations can range from the simple desire to give time and expertise to having a clear objective to use the experience as a stepping stone toward a full-time job or other opportunity. Most situations fall somewhere in between.

It may be helpful to articulate primary and secondary objectives. Perhaps your main objective is to learn about a new industry or gain exposure to a profession related to a career pivot. Perhaps you are just entering the job market and need to acquire a broad range of skills.

For each of these motivations, consider alternative ways to meet your objectives. You could give money instead of time, for example. You could establish discounted rates to serve disadvantaged clients rather than offering purely pro bono services.

Structure and Parameters

There is a significant difference between committing six months to two years as a Peace Corps volunteer, which is a highly structured program that does require the use of relevant expertise; volunteering as an aide in a hospital, transporting patients several days a week for years; and spending a few hours serving food at a soup kitchen once in a while. These represent very different kinds of commitments and fulfill different needs.

Costs, Benefits, and Opportunities

What are the costs of your volunteering or pro bono work to yourself, the organization you are supporting, and the community being served? Consider opportunity costs as well as direct costs. Would the community be better served if you donated money rather than time? From the organization’s perspective, managing volunteers also carries real costs. How are the benefits of your volunteering or pro bono work distributed?

Volunteering and Pro Bono Work Specifically Related to KM

Because KM is not a well-known field across all industries, finding opportunities to volunteer or do pro bono work can be challenging. Organizations with established volunteer programs are unlikely to be seeking KM practitioners and may not recognize the challenges they face as having KM-related solutions, even partial ones.

• KM pro bono work embedded in academic programs. In these cases, the benefits are intended to be clearly educational and experiential for student participants.
• KM pro bono work as a strategy to acquire reputational expertise. This can work for a KM startup in need of testimonials, examples of past work, and potentially sample products that can be shared publicly if planned carefully.
• KM volunteering that emerges from ongoing relationships with organizations, often in the form of “light” KM advice.
• KM volunteering that occurs through engagement in KM communities of practice or other industry-specific communities where KM expertise is shared.
• KM volunteering through mentor–mentee relationships where KM knowledge and experience are passed on.

If you are at a loss and cannot find a satisfying opportunity, and yet you want to acquire KM experience to include on a résumé or LinkedIn profile, consider creating your own personal knowledge management (PKM) project. Identify a KM challenge in your own life.

If your objective is to strengthen your KM credentials, plan to talk about your project publicly, particularly on LinkedIn. Note that discussing KM work exclusively with other KM professionals does not necessarily expand the network of people who need to see your expertise.

Real Example of a Personal KM Project

A 600+ collection of vinyl records. I inherited it. I did not have a clear idea of what was in the collection. I always listened to the same five records, even though I suspected there was much more to discover.

At the same time, I needed a concrete, manageable project to deepen my understanding of taxonomies and ontologies, because reading about them was no longer sufficient.

I developed a digital catalog for the collection, using a formal tagging structure and a tool I was already familiar with.

As a result of this work, which took several months, I gained an ongoing supply of lessons about what I did well, what I learned too late, and a deeper respect for the work of taxonomists and ontologists. I would not be qualified to take on a taxonomist’s or ontologist’s role, but I can confidently say that I could work effectively with one.

As an added bonus, the very real digital catalog and its physical representation have been transformed into almost magical knowledge artifacts in my speculative novel, The Knowledge Cartographer.

Stewardship Starts With Us: Soil, Pesticides, & Litter in our Neighborhood

This is the blog version of a short intervention I was asked to make at our neighborhood association's monthly meeting in April 2025.

1. Introduction

As a volunteer for a local environmental group and an avid cyclist, I see our neighborhood’s landscape up close. The choices we make—whether in our yards, on our streets, or in public spaces—shape the health of our environment. Stewardship isn’t about large-scale changes; it’s about small, intentional actions that, when added together, make a real impact.

2. The Three Issues & Why They Matter

Soil is more than just dirt beneath our feet; it’s a living system that supports plant life, captures carbon, and filters water. Healthy soil means healthier plants, cleaner water, and a more resilient environment. In my own backyard, I work to nurture soil health by composting, using natural amendments, and avoiding synthetic fertilizers. Mulching helps retain moisture and prevent erosion. Simple steps, like leaving leaves to decompose instead of bagging them, skipping chemical-heavy lawn treatments, and planting native species, contribute to healthier soil.

Another concern is pesticides. While they are designed to kill pests, they also harm pollinators, birds, and the microorganisms that keep our soil thriving. These chemicals seep into our groundwater and can impact human health. In our neighborhood, I’ve seen fewer pollinators and dead patches of grass, likely linked to pesticide use. There are alternatives, like organic pest control and integrated pest management, that balance pest control with ecological health. We can also advocate for pesticide-free zones in public areas and talk to neighbors about safer options.

Litter is another issue I notice often, particularly as a cyclist. Trash tends to accumulate in certain areas and finds its way into storm drains, ultimately polluting our waterways. Many of us aren’t the ones littering, but we can still play a role in addressing the problem. Some communities have reduced litter through better-placed bins, community clean-up efforts, and gentle reminders through signage and public engagement. We can encourage businesses to keep their storefronts clean and advocate for policies that reduce single-use plastics. Pairing clean-up events with social gatherings or engaging younger generations can help build a stronger culture of responsibility.

3. Call to Action – How We Can Shape Behaviors

We’re not powerless when it comes to the environment around us. Small, everyday choices create ripple effects. Choosing to compost, skipping a pesticide treatment, or planting pollinator-friendly flowers are all simple changes that, over time, improve our shared space. In our neighborhood, we can support local clean-up efforts or start our own. Leading by example—picking up trash, practicing sustainable gardening, and encouraging others to do the same—helps create a shift in behavior and attitudes.

4. Closing Thought

Every bit of soil we care for, every piece of trash we remove, and every chemical we avoid contributes to the health of our community. If each of us takes small steps, the collective impact will be significant. Stewardship isn’t just about individual action. Stewardship is about fostering a shared commitment to protecting and improving our neighborhood. It's also about making sure we are engaging with the most relevant audiences. The neighborhood association can be a partner in this effort, as a place to ask for ideas to have a much broader reach within the community.

Photo Credit: Barbara Fillip.
Pawpaw flower