Why 70 Pages of Markdown Won't Give You Back Control
An LLM can document your codebase in readable pages you can't verify. A structured spec, linked to the code and checked, answers what they can't.
August 20, 2026 · 3 min read
You asked an LLM to explain your codebase and got back 70 pages of clean Markdown. Everything reads well. And on Monday morning, you are exactly as nervous about touching the code as you were on Friday.
Friday, the pages arrive
The setup is familiar. A team inherits a SaaS product nobody fully understands anymore. Someone points an agent at the repository and asks for complete documentation. Twenty minutes later there is an overview, a file per module, a few diagrams, an endpoint list and a glossary.
People are pleased, and they should be. The writing is clear. Onboarding the next developer just got easier.
Monday, the first real question
Then sales asks for a change. The Team plan should go from 10 seats to 15, and admins should stop counting toward the limit. Simple on a call. In the code, nobody is sure.
So you search the pages for “seat.” Three hits. One says the limit is checked when someone is invited. Another mentions a nightly job that deactivates extra users. A third describes the billing sync. Are these one rule, or three rules that drifted apart over the years? The pages can’t tell you, because they were never built to answer that kind of question.
That is the trouble with generated Markdown, whatever the model. Nothing ties a sentence to the code it describes, so nobody can verify it. When the code changes, the text has no idea. Ask again next week and you get different wording with different gaps. And finding every rule that touches the Subscription object still means reading and cross-checking by hand.
More text. Same fear.
Same question, asked of a spec
Now picture the seat limit as an element of a structured spec instead of a paragraph. It has a type (a business rule), a condition, and links. It points to the invitation screen that enforces it, the nightly job that also enforces it, the Plan object it reads, and the exact files where each check lives.
“What happens if we move to 15 seats and exclude admins?” becomes a query. You might find that the invitation check already excludes admins and the nightly job doesn’t. That is a named gap with two sources attached, which beats a hunch every time.
This is what Lyriks builds. The spec sits in a graph database rather than a RAG pipeline, and the engine that reads it is deterministic, based on graph rewriting and developed in-house. Ask the same question twice and you get the same answer twice.
Before anyone writes code
The seat change can then be simulated against the spec before a line is written. The impact report lists the features, rules, screens and data involved, the conflicts to settle, what to test, and also what stays untouched. In practice that last list is often the one that calms the room. The Evolve use case shows how it works.
Once the change is validated, the spec goes to the coding agent you already use, through MCP, with each rule and its source. Lyriks checks the result rule by rule and sends every gap back until all of them hold. At release, each rule lands in one of four buckets: compliant, added without being asked, forgotten, or regressed.
Keep the pages, change the reference
Nobody needs to delete the generated docs. They are still good for onboarding and for walking someone through a module. They just shouldn’t be the thing you check a change against, because you can’t check code against free text. You check it against rules.
If you want to see the difference on your own seat limits, or on whichever rule makes your team hesitate, a demo is the quickest way to find out.