$ man context-wiki/markdown-as-programming

For decades, programming meant writing code. The programmer decides the logic, writes the implementation, debugs the output. The code is the artifact. In 2026, a second paradigm is emerging. The programmer writes instructions in markdown. The agent writes the code. The markdown file is the artifact. Karpathy demonstrated this with autoresearch, where program.md is the control layer and the agent modifies train.py. Boris Cherny built Claude Code around the same principle, where CLAUDE.md is the project-level instruction file that shapes every agent session.

PATTERN

In autoresearch: the human writes program.md with research directions. The agent modifies train.py, runs experiments, evaluates results. The human never touches Python. In Claude Code: the human writes CLAUDE.md with project rules, workflow instructions, and constraints. The agent modifies the codebase. The human reviews and corrects. Corrections become new rules in CLAUDE.md. In content systems: the human writes voice DNA, anti-slop rules, and platform playbooks. The agent generates content. The human scores and iterates. Quality rules accumulate. In GTM pipelines: the human writes skill files with enrichment patterns, scoring logic, and routing rules. The agent runs the pipeline. The human monitors and adjusts. Lessons accumulate. In every case, the markdown files compound. The code is disposable. The instructions are the asset.

PRO TIP

Markdown is not special as a format. What is special is that it is human-readable, version-controllable, and directly consumable by AI agents. You can write a CLAUDE.md, commit it to Git, track its evolution over time, diff it against previous versions, and the agent reads it on every session start. JSON and YAML configurations do similar things but are harder to read and write. Code comments add context but are scattered across files. Markdown instruction files are centralized, readable, and iterable. The format matches the workflow: write, test, revise, commit.

If the primary programming surface is shifting from code to markdown instructions, the most valuable skill is not writing code. It is writing constraints. Designing the instruction layer that produces reliable, high-quality output from AI agents. This is context engineering. The ability to load the right context, set the right constraints, define the right metrics, and let the agent execute. The code the agent writes is an implementation detail. The context you provide is the architecture. Developers who understand this ship faster. GTM engineers who understand this build systems that compound. The paradigm shift is not coming. It is here.