PlanSpec · Prompt Studio

Tune the extraction agent — prompt, memories & knowledge — toward ~100% accuracy.

◆ GPT-5.5

Match vs actuals

96.4%vs your numbers
Match vs your actuals
96.4%
target 98% · +22.4 pts since baseline
Auto-tune iterations
5
this run · max 8 per run
Discrepancies vs actuals
2
2 over-counted · 0 under · 0 invented
① Configure the agent
② Run & review
③ Auto-tune to target

Extraction prompt edited 2m ago

You are a licensed architect transcribing a 2D floor plan. 1. Transcribe ONLY what is literally drawn — never invent or "improve". 2. One wall per physical wall; don't double-count shared partitions. 3. WIC / linen / closet belong to their room — not separate rooms. 4. A double or sliding door is ONE opening. 5. Anything ambiguous → flags[], never a guess. …
Editable · versioned

Memories 4 active

  • Treat WIC / linen / closet as part of its room — don't count as a separate room.
  • A double or sliding door is one opening, not two.
  • On A / B / C style sheets, extract Style A only unless told otherwise.
  • Garage & mechanical areas: count windows only where a glyph is drawn.

Knowledge base indexed

  • PDFEmpire symbol legend1.2 MB
  • PNGDoor & window glyph reference340 KB
  • PDFSauble plan family — A / B / C2.1 MB

Current results — Sauble 34′

1
AI reads the plan done · 100s
GPT-5.5 transcribed the Sauble 34′ sheet (3 floors).
24 rooms10 doors 17 windows4 stairs
2
Auto-grader reviews it 88% confidence
An automated reviewer checks the extraction before you see it. This is the AI's confidence in its own reading — separate from the match against your numbers below.
88%
Floor 2: 1 extra door flagged (small glyph, low confidence) and 1 unclear room boundary (mechanical vs. laundry). Windows and stairs match the labels.
3
Compare to your numbers
Your numbers — entered by Nader · 2:14 pm
Type the real counts from the plan — the comparison updates as you go.
FieldAIYoursResult
Rooms2423
Doors109
Windows1717
Stairs44
4
Your call
Approve this spec, or send it back with the grader's notes for another pass.

Auto-tune by Claude

Claude rewrites the prompt & memories after each run until you hit your target — or the run budget is spent.

Target accuracy
Max iterations / run
  • Run 174%baseline prompt
  • Run 283%+ memory: don't split closets
  • Run 390%prompt: one wall per physical wall
  • Run 495%+ knowledge: symbol legend indexed
  • Run 596.4%tuned window-count rule · ← current
  • Target98%

Concept — illustrative data. Built on PlanSpec's live extraction engine and human-review workflow. Auto-tune is the next step.