Why the .md File Is the New Agent Programming Language
A few years ago, you programmed an agent with code. Today, you program one with markdown.
This sounds glib until you look at what is actually winning. Karpathy's CLAUDE.md repo is a single markdown file that overrides Claude Code's default behavior across thousands of engineering tasks. Anthropic's Skills format is per-skill markdown. Google's DESIGN.md format (open-sourced May 2026) is design tokens plus voice plus accessibility in markdown. Garry Tan's GBrain uses .md files throughout. Sentie's substrate uses markdown for tenant constitution, skill specs, and playbooks.
Three reasons the pattern keeps winning:
**Foundation models read prose better than code.** A markdown file gives the model intent in language. A JSON config compresses that intent into keys the model has to interpret. Prose wins on edge cases.
**Plain text is reviewable by humans.** Your security team can read a constitution file. Your operations lead can edit it. Your CMO can approve the tone section. No engineering required.
**The spec IS the documentation.** No drift between what the agent does and what the docs say it does. They are the same file.
What Goes in a Real Constitution File
A constitution file is the document the agent reads at the start of every run. Not chat history, not memory - the rules.
The canonical sections, distilled from the multiple instances of this pattern we have seen across vendors:
**Tone and voice.** How the agent talks. "Direct, helpful, no emojis, no exclamation marks." Or whatever fits your brand. This is the equivalent of the brand-voice guide for your communications team.
**Throttles and limits.** Hard ceilings on how much the agent can do per period. "Cap LinkedIn connection requests at 25 per day per inbox." "Cap email sends at 100 per day per sender." "Cap a single Stripe transaction at $5,000 without secondary approval."
**Exclusions.** Who the agent should never contact, what categories it should never address. "Never send outbound to existing customers without an active deal." "Never discuss competitor pricing." "Never offer custom discounts."
**Plan mode default.** "Always show the plan before executing." This is the single most important line for agents that take real action; it prevents the bad-batch failure mode where AI sends 200 messages overnight and gets the account flagged.
**Approval thresholds.** What requires human approval before execution. "Any spend over $500." "Any outbound to enterprise-tier prospects." "Any message in a non-English language."
**Brand-specific rules.** Words and phrases that are on-brand or off-brand. (At Sentie, this includes never using em dashes - that constraint lives in our own constitution.)
**Escalation paths.** Who gets pinged when the agent hits a decision it cannot make alone. "Escalate to account manager for any customer threatening to cancel." "Escalate to success manager for any pricing exception."
**Anti-patterns.** Things that look like good ideas but are not. "Never use 'just checking in' as the opener on a follow-up email - it converts 50% worse than a specific reference." These come from accumulated team learning.
Karpathy's 4 Principles (As a Worked Example)
The CLAUDE.md repo Karpathy released contains four core principles distilled from observing what coding agents get wrong. Paraphrased:
**1. Resist over-engineering.** Choose the simplest solution that works. This overrides the model's tendency to suggest elaborate architectures when a basic one would do.
**2. Do not invent APIs.** If you do not know the SDK, say so and ask. This overrides the model's tendency to write plausible-looking code that fails silently because it invented method signatures.
**3. Do not make up data.** Read the actual file or query the actual database. This overrides the model's tendency to hallucinate plausible values.
**4. Show your work.** Explain what you are about to do before doing it. This is the plan-mode default applied at the per-task level.
These four are the negative space of foundation-model behavior. Karpathy is not telling the model what to do; he is telling it what NOT to do, because the default behavior gets specific things wrong.
A Sentie constitution file does the same thing for your business: encode the not-do patterns specific to your industry, your brand, your customers. The Success Manager builds these out during the assessment phase based on what your team flags.
Why Every Vendor Has One Even If They Do Not Call It That
Some vendors publish their constitution file framework explicitly. Karpathy's CLAUDE.md and Anthropic's Skills are the most visible examples. Sentie publishes the constitution file per-tenant as part of the [Business Brain](/blog/what-is-a-business-brain) substrate.
Other vendors hide it. Salesforce Agentforce, Microsoft Copilot Studio, and most enterprise platforms ship constitution-equivalent rules but they live in proprietary configuration formats inside the platform, not as a reviewable plain-text file the customer can read.
The substantive difference: when the vendor hides the rules, you cannot audit them. You cannot know exactly what the agent will and will not do. You cannot edit a rule without going through the vendor's UI. You cannot version-control the rules separately.
When the vendor publishes the constitution file, your team can read it, edit it, version it, audit it. The agent is governable by your team rather than just operable by your team.
This is a real difference in how much control the customer actually has.
Sentie's Per-Tenant Constitution File
Every Sentie customer gets a constitution file as part of their Business Brain configuration. The Success Manager builds it during the assessment phase based on:
- Your brand voice guide (or, if you do not have one, captured during onboarding) - Your operational rules and never-dos (Sunday job scheduling, no-discount policies, etc.) - Your throttles and approval thresholds - Your escalation map (who gets pinged for which categories of decision) - Your industry-specific compliance requirements
The file is plain markdown. You can read it, your security team can audit it, your CMO can approve the tone section, your operations lead can edit the throttles. Changes ship live to the agent within minutes; the agent reads the constitution on every run.
This is structurally different from a vendor configuration UI. The file is the source of truth, not a screen.
What to Ask a Vendor
Five questions:
**1. Show me the constitution file.** If they cannot or will not, the agent does not have configurable rules - it has hidden defaults.
**2. Can I edit it directly?** Plain-text editing should be available. UI-only editing is a leaky abstraction over the underlying rules.
**3. Can I version-control it?** Git-based version control is the right answer for serious teams. The constitution file is operational policy; it should change with proper review.
**4. How fast do edits take effect?** Minutes is the right answer. Hours or 'next deploy' indicates the file is not really being read on every run.
**5. What is in the default constitution and what is the customer-customizable layer?** Vendors that distinguish 'system defaults' from 'tenant overrides' are being honest about the architecture. Vendors that conflate them are hiding decisions.
At Sentie all five answers are yes-with-detail. Most agent vendors fail at least three of these tests because hidden constitution files are how they maintain control of the agent's behavior.
The Bigger Pattern
The constitution file is one instance of a larger trend: the .md file is becoming the programming language of agents. CLAUDE.md (general behavior), AGENTS.md (cross-vendor IDE agent rules), SKILL.md (per-skill specs), DESIGN.md (design system), memory.md (per-operator context).
If you are evaluating a custom AI agent vendor in 2026, ask which .md files they ship. The answer tells you how AI-native their substrate actually is. Vendors that lead with proprietary configuration UIs are building for the buyer of 2022. Vendors that lead with plain-text spec files are building for the buyer of 2027.
Sentie sits on the second side of that line by design.
[Start a free assessment](/onboarding) to see what your constitution file would look like configured to your specific business. Your Success Manager walks through the sections during discovery.