TRY THE CLM ENCODERS ON YOUR OWN INPUT
The System Prompt Encoder compresses task instructions, role definitions, and operational guidelines into a dense token vocabulary an LLM can still follow.
Task prompts describe a one-off action to run right now ("summarize this transcript"). Configuration prompts define persistent behavior — role, rules, and runtime placeholders like {{customer_name}}. Whether either shape also returns a minimized natural-language remainder alongside the CL token is controlled by OUTPUT MODE below.
FULL returns only the compressed CL token, e.g. [REQ:ANALYZE][TARGET:TRANSCRIPT:DOMAIN=QA][EXTRACT:SENTIMENT,URGENCY]. MINIMIZED appends a trimmed natural-language remainder underneath it, with any role, rule, or output-format text already captured by the CL token stripped out, e.g.
[REQ:ANALYZE][TARGET:TRANSCRIPT:DOMAIN=QA][EXTRACT:SENTIMENT,URGENCY] Return the analysis as JSON with sentiment and urgency fields.
See Configuration Prompt for the full minimization walkthrough, including placeholder binding.
If the compressed form (CL token, plus the minimized remainder when applicable) would not be smaller than the original prompt — i.e. the compression ratio is not positive — CLM returns the original input unchanged instead. See CLM Output for details.