Prompt Compass in action: before and after prompt improvement from a real session
Prompt Compass in action
Real session · Advanced mode · Human in the loop
How a voice-memo brain dump became a production-ready AI memory policy
Before · Raw voice memo
like, um, going forward, we create a protocol to separate any of the personal overlay space from, like, my personal prompt compass MCP is stored in it, the writeables that cloud can write to, away from the rest of the MCPs, so that that way everybody who downloads it gets a blank slate with no personalization yet...
Filler words, doubled phrases, misspelled tool names — model has to guess intent
No role defined — model doesn't know what posture to take
No output contract — what should be produced? A doc? A checklist? A plan?
Deferred items tangled with must-haves — model can't tell what to do now vs. later
Pattern: Voice Memo / Dictation Intake · Anti-patterns: hallucination-bait, scope sprawl, missing success test
3 Q&A
stages
stages
After · Prompt Compass improved
Role: Technical packaging assistant
Objective: Sequenced packaging checklist for shipping Prompt Compass as a distributable product.
Must-haves (in order):
1. Clone + sanity check — confirm personal overlay and prompt library are cleared.
2. .MCPB packaging — bundle MCP into .mcpb file.
3. Raw zip packaging — note file exclusions.
4. agents.md file — define structure.
Exclusions: Skill (deferred). Plugin (deferred).
Output Contract: Numbered checklist. Flag deferred items [DEFERRED]. Success test: a dev could hand this to a collaborator and they'd know exactly what to do.
Role anchor — model knows exactly what posture to take
Priority ordering extracted from the stream-of-consciousness — 4 deliverables, numbered
Deferred items explicitly flagged [DEFERRED] — model skips them cleanly
Output contract with success test — model knows when "done" is done
RACCCA score: 21 → 26 · All 3 anti-patterns resolved
What happened in the middle — the HITL loop
1
Prompt Compass detected the pattern — Voice Memo / Dictation Intake. Diagnosed 3 anti-patterns before rewriting anything.
2
Stage 1 — Intent & contract. Asked: who is the audience? Kyle answered: "All of the above — Claude, Codex, developers, Kyle directly." This changed the depth and tone of the output contract.
3
Kyle pivoted the objective mid-loop. Original brief said "write at session end." Kyle answered: "I jump between conversations all the time — session end is meaningless." Prompt Compass rewrote the core policy around event-based writing instead. This could not have been inferred from the original prompt.
4
Stage 2 — Context grounding. MCP tool names named explicitly. Floor/ceiling authority model locked in. Codex's critique incorporated. Two output variants produced: terse and full.
+5
RACCCA score lift
3
anti-patterns resolved
3
HITL questions asked