adaptive-reasoning

The whole skill is a scoring rubric. Add weight for multi-step logic, ambiguity, architecture, formal reasoning, novelty and stakes; subtract for routine work, clear single answers and simple lookups. Two thresholds matter in practice, and the subtractive terms are the underrated half. Without them almost every request looks moderately complex and the agent reasons about everything, which is the exact failure mode the rubric exists to prevent. The instruction to activate without asking is also right for a pre-processing step, since a prompt asking permission to think would cost more than the thinking it saves. Scores are self-assessed, so treat the rubric as a discipline for consistency rather than a measurement.

Use case: Used the rubric to decide which scheduled requests were routine enough to answer directly and which deserved extended thinking.

# Calibrating the reasoning threshold Adaptive reasoning scores each incoming request and picks a reasoning level from the score. The scoring table is the easy part; the calibration is where it either helps or turns into constant deliberation. ## The score Positive signals: multi-step logic, ambiguity, code architecture, formal reasoning, novelty and stakes. Negative signals: routine or repetitive work, a single clear answer, and simple lookups. The negative terms carry the same weight as the positive ones, and they are the half most people forget. A rough implementation to keep the arithmetic honest: ```python positive = {"multi_step": 3, "ambiguity": 2, "architecture": 2, "formal": 2, "novel": 1, "high_stakes": 1} negative = {"routine": -2, "single_answer": -2, "lookup": -3} def score(flags): return sum(positive.get(f, 0) for f in flags) + sum(negative.get(f, 0) for f in flags) def action(total): if total <= 2: return "fast" if total <= 5: return "standard" if total <= 7: return "deliberate" return "extended" ``` ## Calibrating on a real workload Run the rubric over a week of real requests and count how often each band fires. If almost everything lands in standard, the negative terms are not being applied. If everything lands in extended, the positive terms are being over-applied and the agent is paying for thinking it does not need. ## Why not ask first The skill says to activate without asking. That is correct for a pre-processing step: a prompt asking permission to think costs a round trip and more tokens than the thinking it was meant to save. ## Tips - Apply the negative signals deliberately; without them nearly every request looks moderately complex. - Gotcha: the score is self-assessed, so it drifts optimistic under pressure. Spot-check it against real workloads instead of trusting it. - Treat the threshold as a nudge toward consistency, not a measurement of difficulty. - Recalibrate after any change to the agent's model or prompt; the scoring rubric was tuned for the old behaviour. - Keep the fast band genuinely fast; if simple lookups start reasoning, the gate has stopped working.