Entering the Markdowns Era

Why prompts feel flat

Is it just me, or does the era of prompts feel flat? The output is not wrong, and it is not exactly disappointing. It is just flat, and I keep wondering what comes next.

In 2024 and 2025, I learned how to prompt. I experimented with frameworks that front-load context, instruct AI to ask clarifying questions, play a particular role ("you're a finance expert with 20 years experience"), and shape the output I want. For many people, that kind of practice produced a real productivity gain—often on the order of 10 to 30%, depending on role and task.

Even so, AI can still feel like an extra limb: powerful, but not quite woven into how I work. It makes me faster without always feeling transformational, and it rarely pulls every moving part into the same frame the way I know it could.

So the more interesting question, for me, is not how to write a better prompt in the moment. It is how to work with AI in a way that feels natural, stays consistent, and scales without constant reinvention.

My stance on prompts

To be clear, this is not an argument against prompts. They are useful, and they are often how people first see what AI can do.

But they run into structural limits when I try to move past one-off use or string work together across many sessions. The first limit is inconsistency. A model can only work from the context you give it, and that context depends on what someone thought to include in the moment. Nuance can be missing, out of date, or framed a different way the next time around.

The second limit is variability across people. Everyone uses slightly different wording and carries different unstated assumptions into the chat, so two colleagues can ask what is essentially the same question and still get answers that do not quite line up. When I am working alone, I can live with that drift. When several people need one shared basis for a decision, it becomes a problem: each prompt is its own closed episode, while the work I care about accumulates over time and across roles.

From prompting to persistent context

That is where markdown files come in. I was recently introduced to the idea of using them as persistent sources of truth: structured documents that hold the current thinking on a topic and can be attached to a conversation so the model can help with analysis, insight, and execution.

The closest analogy I have is a company policy—a Delegation of Authority, a procurement standard, something like that. Unlike a policy written to sit in a folder, a markdown is built to change. It is written in plain language, updated as understanding moves, and meant to be used in day-to-day work with AI.

Markdowns do not replace prompts; they change what a prompt is for. Instead of using every session to rebuild context from memory, I can point the model at a shared file and use the prompt to give instructions on top of that foundation.

Using AI to create the markdowns themselves

One part of this is easy to skip: I did not need perfect markdowns on day one. In practice, AI has been very good at helping me create them.

I still use prompts to turn messy material into something structured. I might feed in meeting notes, a policy draft, a strategy deck, or a page of rough bullets and ask for a clear, well-structured markdown file. I then edit, tighten, and—when the work is shared—align with colleagues on what belongs in the final document.

Prompts are how those files get started. Over time, the outputs from early prompting settle into markdowns that stay reusable: shared context I do not have to recreate every time.

In that sense, prompts are scaffolding. The markdown is what remains standing.

Thanks for your help! Can you save this as a markdown file that I can download?

Agent

Sure, here's your .md file to download!

Example: Large corporate transformation initiative

Picture a transformation team running several workstreams at once. Instead of re-explaining the same background on every call, they attach a handful of markdown files and ask the model to work inside that shared frame.

With those markdowns attached, the model already understands objectives, governance, and budget. The prompt is no longer where memory lives; it is the lever that kicks off analysis.

Example: Startup launch

The same pattern holds for a startup, where speed and cross-functional alignment can matter as much as depth of analysis.

That is how a small group can move quickly without losing a coherent story.

Early guidelines on an emerging practice

This way of working is still forming, but a few habits already stand out.

Less is more. A small set of strong markdowns beats a pile of overlapping ones. Each file should own a clear scope and serve as the single place to look for that topic—much like roles on a team, where overlap breeds confusion and gaps leave exposure.

Keep them concise. Concise does not mean shallow. Length is fine when it earns its place; padding is what hurts. Someone still needs to own clarity and trim what drifts.

Plan to revise. Markdowns should change as understanding deepens—edited, tightened, sometimes restructured. That maintenance is not bureaucracy; it is how durable knowledge compounds.

Natural language meets computer speak

Markdown sits in a useful middle ground: human language with enough structure for machines. Next to slides or PDFs, models spend less energy fighting the container and more on the substance.

Less translation. More thinking.

# Transformation Program Overview

## Purpose
- Reduce operating cost while improving delivery speed
- Establish repeatable decision-making frameworks

### Decision rule
- Prioritize **customer value** over `local optimization`.

## Principles
1. Decisions favor long-term value over short-term optics
2. Speed matters, but clarity matters more

## Key Dates
- Program kickoff: March 1, 2026
- Steering Committee: Monthly
- Target completion: December 2026

Embedding AI into the core analytics engine

Where this pays off most is coherence at scale.

When I share markdowns across a team, AI stops being only a personal shortcut and starts acting like an analytical partner working from the same brief. Different people can stress-test scenarios or compare notes, and the answers stay tethered to one shared context.

Compared with prompting from a blank slate each time, the experience feels less like shouting questions into a void and more like working beside someone who read the same packet first.

I have watched small teams punch above their weight this way—not because the model is magic, but because the context finally travels with the work. That is when AI moves from feeling like an extra limb to feeling like part of how the team operates.

Conclusion

There is still a great deal to learn.

Markdowns are not a silver bullet. They take discipline to maintain. Even so, they mark a shift—from one-off prompting toward understanding that persists and can be shared.

Prompts helped me discover what AI can do. Markdowns help AI understand how I work.

That feels like a solid place to keep building.