How this channel uses the term Prompt debt appears in Jared Zoneraich's discussion of what happens when a prompt is patched case by case instead of designed: edge-case fixes keep accreting until the system prompt becomes a long run-on list of special instructions. [^d351f8cb-2125-5ece-8488-6f6ee66c9ee2@2552] The evidence: leaked system prompts Zoneraich points to the leaked ChatGPT and Claude system prompts as the visible example. He reads their length and patchwork structure not as thoroughness but as overfitting to no specific use case — the prompt has absorbed so many one-off fixes that it no longer serves any single purpose well. [^d351f8cb-2125-5ece-8488-6f6ee66c9ee2@2552] The alternative he advocates is a workflow — a DAG/graph — that routes each request to a small prompt that does one thing really well: discrete prompts are easier to test, easier to collaborate on, and fail less. [^d351f8cb-2125-5ece-8488-6f6ee66c9ee2@1632] Why it matters - Prompt debt is a smell, not a milestone: a growing system prompt usually means patches are accumulating faster than the design is being cleaned up. - It contrasts directly with the routing approach — many small, single-purpose prompts instead of one ever-growing mega-prompt. - When you spot edge-case patches stacking up, treat it as a signal to restructure rather than add another clause. [^d351f8cb-2125-5ece-8488-6f6ee66c9ee2@1632] [^d351f8cb-2125-5ece-8488-6f6ee66c9ee2@2552]