Miles prompt compression
I compressed the measured Miles prompt across two passes—from 4,892 tokens to 3,732. The recordings below are the latest Pass 02 pair: 4,146 → 3,732 tokens, with Happy Path still at 9/9.
Shorter prompts are easy. Shorter prompts that still do the job need evidence.
This Dry Dock entry shows how I compressed Miles into Miles Junior across two measured passes, protected the working call behaviours, and checked the result against a defined Happy Path.
Before
This recording is the Pass 02 start state—the Pass 01 Miles Junior result before further compression. The measured prompt was 4,146 tokens. The reviewed call still passed all nine Happy Path checks, including the CRM lookup tool corrected in Pass 01.
Stage proof
This recording is the Pass 02 start state—the Pass 01 Miles Junior result before further compression. The measured prompt was 4,146 tokens. The reviewed call still passed all nine Happy Path checks, including the CRM lookup tool corrected in Pass 01.
Before
Hear the Pass 02 start state before the second compression pass.
Pass 02 before sample — recorded call.
Process
Each pass reduced repeated prompt structure while keeping the call safeguards, then compared one before call with one after call on the same nine-point Happy Path checklist. Pass 02 starts from the Pass 01 Miles Junior result, not from the original Miles baseline.
Step 1
Pass 01 — 4,892 → 4,146 tokens (−746 / 15.25%); Happy Path 8/9 → 9/9; CRM lookup tool corrected.
Step 2
Pass 02 — 4,146 → 3,732 tokens (−414 / 9.99%); Happy Path stayed 9/9 on both reviewed calls.
Step 3
Each pass rechecked behaviour, tool use, timing, and cost against the same checklist—without treating N=1 runtime moves as proof.
Compression passes
This walkthrough shows how the compression work is done. It stands alone as process proof; it is not tied to a single pass.
Process evidence
A standalone walkthrough of the prompt-compression process used on Miles / Miles Junior—how the work is measured, reduced, and retested.
Fallback path
If the embed is blocked or skipped, use the written notes below or open the video directly on YouTube.
Open on YouTubeTranscript notes
YouTube evidenceNote 1
Shows the compression method: measure the prompt, reduce repetition, keep call safeguards, and retest the Happy Path.
Note 2
Supports the Process tile as method evidence; it is not a substitute for the written token and checklist results.
Note 3
Read the Evidence note for N=1 limits and unverified deployment relationships.
Prompt diff notes
After
After Pass 02, the measured prompt is 3,732 tokens (−414 / 9.99%). The reviewed Happy Path stayed at nine of nine. Across both passes from the original Miles baseline to current Miles Junior, that is 1,160 fewer tokens (−23.71%).
Stage proof
After Pass 02, the measured prompt is 3,732 tokens (−414 / 9.99%). The reviewed Happy Path stayed at nine of nine. Across both passes from the original Miles baseline to current Miles Junior, that is 1,160 fewer tokens (−23.71%).
After
Hear the Pass 02 result after the second compression pass.
Pass 02 after sample — recorded call.
What this proves
Good Voice AI work protects the conversation while changing the system underneath it—and shows clearly where the evidence stops.
Signal 01
The behaviours that already passed on Miles stayed intact across both passes, including message capture, number confirmation, the exact closing phrase, and one-question-per-turn discipline.
Signal 02
The review checked tool selection and call behaviour, not just whether the conversation sounded plausible. Pass 01 corrected the CRM lookup tool Miles missed; Pass 02 kept that check green while cutting more tokens.
Signal 03
Individual samples still show mixed timing and cost movement. I am not presenting prompt reduction as a runtime-performance improvement.
Chronology
Dated entries record the reviewed compression comparisons behind this page.
2026-07-13
Issue
The Miles baseline prompt measured 4,892 tokens and the reviewed call missed the required CRM tool-selection check (8/9).
Process applied
Patrick reduced repeated prompt instructions and compared one before call with one after call against the same nine Happy Path checks.
Outcome
The post prompt measured 4,146 tokens (−15.25%). The reviewed after call passed all nine checks, with N=1 and deployment-verification limits.
2026-07-16
Issue
The Pass 01 result still measured 4,146 tokens; the next reduction needed another measured pass without losing the 9/9 Happy Path.
Process applied
Patrick tightened remaining duplication in the Miles Junior prompt and compared one before call with one after call on the same checklist.
Outcome
The post prompt measured 3,732 tokens (−414 / 9.99%). Both reviewed calls passed all nine checks. Cumulative Miles → current Junior: −1,160 tokens (−23.71%).
Next step
If your voice agent has become difficult to maintain, I can help you reduce the prompt while protecting the call behaviours that matter.
We start with a real call path, define what must keep working, make the prompt change, and compare the result. You see both the improvement and the limitations before deciding what ships.
Review your agent
