ecdsafail_skills

Use this skill whenever designing, reviewing, or modifying ECDSA Fail / quantum ECC point-addition circuits to reduce qubit count, Toffoli count, or score. This skill is especially relevant when comparing SOTA submissions, borrowing ideas from Andre/Schrottenloher-style space-optimized arithmetic, Trail of Bits trailmix-style reversible-circuit insights, or choosing structural optimization knobs before island hunting.

Use this skill whenever working on ECDSA Fail / EC point-addition island hunting, GPU nonce scans, clean island validation, cls/pha/anc triage, random-nonce repair readiness, SOTA score comparison, or deciding whether a quantum circuit optimization is worth scanning. This skill is especially important when the user asks to start, monitor, triage, validate, or extend island scans; compare against SOTA; run fast/full validation; repair dirty candidates before scanning; or choose between circuit routes.

Reduce the peak (maximum simultaneous) qubit count — the "width" — of a reversible or quantum circuit. Use this whenever a circuit uses too many qubits/ancilla/scratch, when optimizing circuit width or the qubits-times-gates score, when an arithmetic or Shor-style circuit needs to fit a smaller machine, or when someone asks how to "shave qubits", "lower the qubit count", "reduce ancilla", "free scratch earlier", or "trade gates for qubits". Trigger even if the user describes the goal informally (e.g. "this adder needs way fewer wires", "where's all my space going", "can I reuse that register") without naming peak width. Provides a left-to-right timeline analysis: find the tallest moment, inventory the qubits alive there, identify whether the peak is a single binder or a multi-phase plateau, and brainstorm uncompute-early/recompute-later holes, live-range shortening, truncation, transcript compression, and coordinated co-binder reductions to cut it.

Reduce the Toffoli / T-gate / non-Clifford count of a reversible or quantum circuit — the expensive, magic-state-bound resource that dominates error-corrected runtime. Use this whenever someone wants to "cut the T-count", "lower the Toffoli count", "reduce non-Clifford gates", "need fewer magic states", "my carry uncomputation is half my gates", or improve the gate factor of a qubits×Toffoli score. Covers venting carry-uncompute into measurement+phase fixups, conditional/measured replay so gates only execute on the branch that needs them, fused adders/folds, cross-block algebraic fusion of modular add/negate chains, ancilla-free majority, skipping provably-redundant cleanup, and avoiding full materialization. Fires on informal phrasing ("too many Toffolis", "T-count is my bottleneck", "can I make this uncompute free") and distinguishes emitted vs average-executed gate counts. For lowering the QUBIT/width count instead, use peak-qubit-reduction; this skill owns the GATE axis.

Use this skill when coordinating two or more agents, such as Claude and Codex, on the ECDSA Fail challenge or another benchmark-driven circuit-optimization hunt — or when one agent fans out parallel sub-investigations (multi-agent autoresearch). Trigger whenever the user mentions collaborating with another agent, a mailbox, Claude/Codex teamwork, exchanging hypotheses, independent cross-checks, splitting GPU/research work, fanning out parallel research/diagnostic subagents, or converging on a route to beat SOTA. This skill turns multi-agent chat into a disciplined scientific loop: read the thread, share exact evidence, fan out orthogonal sub-investigations, divide experiments, reconcile convergence/conflict, and preserve the winning route.

Validate and debug the CORRECTNESS of a reversible or quantum circuit — not its size. Use this whenever a circuit "gives wrong answers on some inputs", "ancilla aren't returning to zero / uncomputation is leaving garbage", there are "phase errors", "it passes my basis-state test but I'm not sure it's right in superposition", or someone needs to test a reversible adder/multiplier/modular-arithmetic block before trusting it inside Shor's. Covers the classical/phase/ancilla (cls/pha/anc) three-channel decomposition, differential testing against a trusted reference on RANDOM inputs, the forward-inverse identity check, telling a designed approximation failure apart from a real bug, and localizing the first broken op. Fires on informal phrasing ("my uncompute is dirty", "results get corrupted when I compose it", "is 0.3% failure expected?"). For making a correct circuit SMALLER use peak-qubit-reduction (qubits) or toffoli-reduction (gates); this skill owns CORRECTNESS.

Reverse-engineer a competitive optimization benchmark's advancing frontier: when a competitor or new SOTA submission beats your score and the winning entries are visible (a git history of accepted submissions, public leaderboard solutions, released configs), extract exactly which lever drove the improvement and whether it transfers to you. Use whenever "a new SOTA was submitted, analyze it", "diff the new leaderboard winner", "what did they change / what lever did they use", "reverse-engineer their improvement", "figure out what drove the score jump", or you need to learn from a competitor's accepted entry on a benchmark you're optimizing. Covers diffing consecutive winners, walking every intermediate submission (not just endpoints), re-measuring the metrics yourself, decomposing a multi-factor score delta, separating a transferable lever from a non-transferable search artifact, and tracking the moving bottleneck so you know the next target. For DOING the optimization, use peak-qubit-reduction / toffoli-reduc