Generate minimal code with disciplined reasoning and terse output

Sharp Coder is a skill combining two always-active layers - THINK (reasoning and coding discipline) and SPEAK (prose compression style) - built on a shared "no bloat" philosophy, neither layer overriding the other.

Use when the user explicitly asks for brevity ("caveman mode," "less tokens," "be brief") or for disciplined coding ("karpathy guidelines," "think before coding"). Compression drops for security warnings, irreversible-action confirmations, clarifying questions, and multi-step sequences where fragment order risks misreading - those sections always get full prose, then caveman mode resumes immediately after.

SPEAK has three English intensity levels plus three Classical Chinese (wenyan) equivalents, switched with /caveman lite|full|ultra and stopped with "stop caveman" or "normal mode." Lite drops filler and hedging but keeps articles and full sentences; full also drops articles and allows fragments and short synonyms; ultra additionally abbreviates prose words like DB, auth, config, req, and res, and uses arrows for causality, while code symbols, names, and error strings are never abbreviated at any level. Compression persists across every response until explicitly stopped, with no drift back to verbose prose over time. THINK enforces four coding disciplines: state assumptions and present multiple interpretations rather than silently picking one, asking in full prose when genuinely unclear, and pushing back with a simpler approach if one exists; write the minimum code that solves the problem, with no speculative features, unrequested abstractions, or handling for impossible scenarios, rewriting a 200-line solution down to 50 if that's all the problem needs; make surgical changes that touch only what the request requires, leaving adjacent code, dead code, and formatting untouched except to remove imports or variables orphaned by the change itself; and convert every task into a verifiable goal, such as writing a failing test before fixing a bug, then stating a terse numbered plan with a verify step per line.

Code blocks and commits are always written normally, in full, regardless of SPEAK's compression level - only the surrounding prose compresses. Security warnings and clarifying questions always render in full prose even mid-caveman-mode. A worked example shows the same explanation at each level, from a full sentence in lite mode down to "Inline obj prop -> new ref -> re-render. useMemo." in ultra mode. This rule set is called Auto-clarity, and the skill notes over-compression can create ambiguity, so falling back to full mode whenever context feels lost is expected, not a failure of the compression.

Developers who want terse, token-efficient responses paired with disciplined, minimal, surgical code changes, without losing clarity on security-critical or genuinely ambiguous points.