Example packs · PromptParle

Savings depend on the dial and how noisy the pack is. We are still measuring real workloads; published examples (Noisy ~78%, Security ~60%, Clean ~2%) show the shape of results so far. Clean unique prose often barely moves. Noisy logs, security packs, multi-file dumps, and agent chains are where reduction usually shows up. Proof lives in the dial savings line in the desktop UI per turn.

Numbers below are illustrative demo passes, the kind of work where the dial earns its keep (or correctly does almost nothing).

Noisy OT / app log

A long log dump with timestamps, repeated health checks, stack noise, and one real failure buried in the middle.

Profile: log-analysis · Dial 4

−78%

18,400 → 4,120 tok

~14,280 saved

Logs are full of repeated lines, boilerplate headers, and low-signal chatter. The optimizer keeps the error neighborhood and unique events; it drops redundant heartbeats and duplicate frames.

Before (shape)

[2026-07-11 08:01:02] INFO  sensor=CVIL-A  heartbeat ok latency=12ms
[2026-07-11 08:01:07] INFO  sensor=CVIL-A  heartbeat ok latency=11ms
[2026-07-11 08:01:12] INFO  sensor=CVIL-A  heartbeat ok latency=12ms
… (120 similar heartbeats) …
[2026-07-11 08:14:33] WARN  plc=PLC-7  modbus timeout unit=3 fn=0x03
[2026-07-11 08:14:33] ERROR edge-agent  reconnect attempt 1/5
[2026-07-11 08:14:34] ERROR edge-agent  reconnect attempt 2/5
[2026-07-11 08:14:35] INFO  sensor=CVIL-A  heartbeat ok latency=14ms
… (more heartbeats + full stack traces repeated thrice) …

After (shape)

Log focus (dial 4 · log-analysis)
• WARN plc=PLC-7 modbus timeout unit=3 fn=0x03 @ 08:14:33
• ERROR edge-agent reconnect 1/5 then 2/5 immediately after
• Heartbeats: ~120 identical CVIL-A ok lines collapsed (latency ~11-14ms)
• Repeated stack frames deduped; unique frames kept
Ask: root cause and next check on PLC-7 path.

Security review pack

A security pass over multiple files: auth helpers, config, and a long dependency list pasted “just in case.”

Profile: security-review · Dial 3

−60%

12,600 → 5,040 tok

~7,560 saved

Security reviews need auth boundaries, trust edges, and risky APIs, not every import line and license banner. The profile biases toward secrets patterns, auth paths, and network egress; filler docs and duplicate configs compress hard.

Before (shape)

// package-lock excerpt (2k lines)
// LICENSE × 4 copies
// README install notes…
export function verifySession(token: string) {
  // … full 80-line implementation …
}
export function requireAdmin(user: User) { … }
// docker-compose with every service commented twice
// full .env.example with 40 unused keys

After (shape)

Security-review pack (dial 3)
Keep:
• verifySession / requireAdmin control flow + failure paths
• Trust boundary: session cookie → API key → provider call
• Egress: provider HTTPS only; desktop tools stay local
• Secret-shaped config keys (masked if present)
Collapsed: lockfile noise, repeated LICENSE/README, duplicate compose comments
Ask: auth bypass, privilege gap, secret leakage in this pack.

Clean unique prose

A short, carefully written product question with almost no repetition, already dense.

Profile: general · Dial 3

−2%

420 → 410 tok

~10 saved

There is little bloat to remove. Unique prose is already signal. Expect near-zero reduction, and that is correct behavior, not a failure.

Before (shape)

We're evaluating PromptParle for a small security team. We use Claude and Grok with our own keys. Can we keep flagship models while cutting log-heavy context, and does anything store our prompts?

After (shape)

We're evaluating PromptParle for a small security team. We use Claude and Grok with our own keys. Can we keep flagship models while cutting log-heavy context, and does anything store our prompts?
(≈ same size, already compact)

Attached product user guide

Attach a long product user guide and ask for an executive summary. Local prep keeps signal; dial 3 cuts bulk before Grok sees the turn.

Profile: executive-summary · Dial 3

−86%

100,000 → 14,000 tok

~86,000 saved

Long guides carry repeated structure, boilerplate, and low-signal chrome. The desktop client optimizes on the PC, then the model writes the summary. Live UI showed −86% tokens and a downloadable executive summary — one real turn, not a lab mock.

Before (shape)

[ATTACH] Product User Guide (full document)
===== FILE: Product_User_Guide.pdf =====
… hundreds of pages of product UI, procedures, screenshots captions,
repeated section headers, license-ish banners, and operational detail …

User: Provide me an executive summary of the Product User Guide

After (shape)

Context after dial 3 local prep (shape):
• [ATTACH] primary this-turn file kept as evidence
• Boilerplate / near-dupes collapsed; hot sections retained
• ~100k → ~14k tokens (−86%) before grok-4.5
Deliverable: Product_User_Guide_Executive_Summary.md
(Real desktop savings line; results vary by document.)

Published example packs

Noisy OT / app log

Security review pack

Clean unique prose

Attached product user guide

When savings are high vs near-zero