skill-zerolisted

# Skill ZERO Two-mode skill. The **skeleton** is the GPT-5.6 Sol execution framework (general task execution). The **meat** is a domain specialization for coding-agent interpretability via linear probes (Silva, Tu, Monperrus 2026). When the task lands in the coding-agent-interpretability domain, switch from general mode to specialist mode while keeping the general framework as the outer loop. > Parent skills: `think-like-gpt-5-6` (general framework), > `coding-agent-interpretability` (domain specialization). > This skill is a synthesis — both parent skills remain installed for > their own direct triggers. Use this skill when you want both at once > OR when you want to add a new domain specialization to the pattern. ## Inputs to collect In all cases, the intake record from `references/principles-and-loop.md` §1.1: - Objective, deliverable, audience, constraints - Must-preserve, prohibitions, evidence requirement - Deadline relevance, tool need, risk level, acceptance criteria - **Domain classification** (NEW for Skill ZERO): - General mode (default) — most tasks - Coding-agent-interpretability mode — task mentions hidden states, residual streams, latent program representations, latent programming horizon, edit-outcome prediction, mechanistic interpretability of coding agents, or references arXiv:2607.05188 Skip intake for trivial (complexity ≤ 4) tasks. ## The 7-step core loop (apply on every non-trivial task) > **Observe → Interpret → Decide → Act → Ve