simplest-probe-first

Simplest-Probe-First

A tunnel-visioned agent commits to the first (often heaviest) approach and never checks the cheap one. Same-context self-talk ("maybe consider a simpler option") is unreliable (intrinsic self-correction degrades reasoning — Huang et al., ICLR 2024). The reliable fix is a deterministic probe: run a one-command external check whose result picks the path for you.

The rule

Before any approach that costs more than a few minutes or spawns agents: name the cheapest falsifiable probe, RUN it, and let its result choose the path. If the probe says the simple path works, take it. Never skip the probe because the hard path "feels right."

Probe table (extend it as you learn)

Situation	Cheap probe to run FIRST	If the probe says…
PDF / scanned doc → extract text	pdffonts file.pdf or pdftotext file.pdf -	empty / no fonts = no text layer → use OCR (ABBYY, the ocr MCP, tesseract), NOT a vision-LLM pipeline
Office / HTML doc → text	check for a text layer / try markitdown	has text → parse directly, skip OCR and vision
Reverse-engineer a function	static decompile (IDA / Ghidra / Binary Ninja) + xrefs first	static answers it → done; only go live-hook / Frida if static is genuinely insufficient
"Make X faster"	profile it once	the profile names the hot path → fix that; do not rewrite blindly
"Build utility Y"	gh search code + search npm / PyPI / crates	it exists → fork / port / wrap (see the development-workflow rule), do not hand-roll
"Have I solved this before?"	mcp__basic-memory__search_notes + grep ~/.claude/*.md + project docs	found → reuse the documented approach, do not re-crack it
Big new abstraction / orchestration	state the complexity tier and why a simpler tier fails	can't justify it → use the simpler tier (anti-over-engineering)

Output

State, in one or two lines: the probe you ran, its result, and the path you chose because of that result. Then proceed.