thalarch-reasoninglisted

Adaptive deliberation layer for difficult engineering and design work. Use when a task is ambiguous, high-risk, multi-step, architecture-heavy, debugging-heavy, version-sensitive, cross-module, or otherwise likely to reward slower reasoning. Forces explicit problem framing, competing hypotheses/approaches, disconfirmation, independent challenge, evidence-based adjudication, uncertainty tracking, and a final falsifiable proof without exposing private chain-of-thought or adding ceremony to trivial work.
LUC4N3X/antigravity-thalarch · ★ 2 · Code & Development · score 65

Install: claude install-skill LUC4N3X/antigravity-thalarch

# Thalarch Reasoning Harness The goal is not to imitate another model's writing style. The goal is to make the current model behave more deliberately: delay commitment, separate facts from guesses, generate real alternatives, seek disconfirming evidence, use independent contexts when useful, and stop only when the conclusion is supported by evidence appropriate to the claim. Do **not** expose private chain-of-thought. Externalize only compact decision artifacts: facts, unknowns, candidate explanations, evidence, rejected alternatives, decisions, risks, and proofs. ## 1. Choose deliberation depth Pick the smallest depth that fits the task. ### D0 — Direct Use for trivial, deterministic, reversible edits with obvious scope. - understand the exact request; - make the smallest change; - run the nearest check; - stop. No hypothesis theatre and no multi-agent ceremony. ### D1 — Guarded Use for small but non-trivial work. - frame the contract; - identify the main assumption that could be wrong; - inspect the relevant evidence; - implement; - verify. ### D2 — Deliberate Default for meaningful debugging, feature, refactor, API/data, or design work. Before commitment: 1. build a compact problem representation; 2. separate `FACT`, `INFERENCE`, and `UNKNOWN`; 3. generate at least two plausible explanations/approaches **when the problem genuinely has alternatives**; 4. state what evidence would falsify the leading candidate; 5. gather discriminating evidence; 6. commit o