prove

Explore options and record decisions via structured brainstorming

Structured debugging protocol — sequential hypothesis testing with backtracking

Prune stale cached prove plugin versions from Claude Code's plugin cache

Anchor prove context before a /compact and rehydrate after — survive a context squeeze without losing claude-prove state

Socratic quiz on recent diffs to build deep comprehension of agent-generated code

Stage-2 prose-quality judge for the reasoning-log Review Brief. Judges whether a synthesized brief's narrative is accurate, risk-forward, and coherent against the run's reasoning log. Read-only, advisory, non-blocking — preservation is proven mechanically by `acb brief validate` (Stage-1), so this judges prose only. Invoked by the `reasoning-brief` skill.

Post-agent code quality auditor. Deep codebase audit with surgical refactoring after parallel agent workflows.

Prompt optimization specialist. Audits and rewrites prompts, agent definitions, system instructions, and CLAUDE.md files for LLM efficiency. Use when writing, reviewing, or improving any prompt or agent definition. Triggers on prompt optimization, token efficiency, instruction tuning, or LLM best practices.

Memory-layer curator for prove teams. Reads one team's full memory (Lore, annotations, contributor artifacts) or the project Codex (scrum decisions) and returns a keep/consolidate/promote/supersede plan that maximizes future accuracy per token. Read-only and plan-only — the janitor skill executes every write through the CLI. Invoked by the `janitor` skill with prepared inventory dumps; not designed for ad-hoc use.

Reference for steward skills orchestrating the Progressive Context Distillation pipeline.

Interactive brainstorming sessions for software architecture, product scoping, and general engineering. Use when the user wants to explore ideas, gather requirements, narrow down solutions, weigh trade-offs, or make technical decisions. Triggers on "brainstorm", "let's think through", "help me decide", "what approach should", "pros and cons", or any open-ended design/architecture discussion. Saves decisions to .prove/decisions/ directory.

Semantic commit assistant. Reads scopes from .claude/.prove.json, detects scope gaps and offers to register new ones, groups changes into logical units, and creates conventional commits.

Post-diff Socratic quiz that builds deep comprehension of agent-generated code. Analyzes recent changes, generates causal/design questions, quizzes the developer interactively, and logs comprehension gaps.

Create Claude Code skills, slash commands, subagents, or technical specs. Dispatches by type. Use when the user wants to create a new skill, new slash command, new subagent, new agent, write a spec, draft a specification, formalize a design, audit a spec, promote a decision record, or needs help with skill descriptions, command frontmatter, agent tool permissions, or RFC-style technical documentation. Triggers on "create a skill", "new skill", "create a command", "new slash command", "create an agent", "new subagent", "write a spec", "spec for", "draft an RFC", "formalize this", "audit this spec", "revise the spec".

Lift reasoning-log findings into durable scrum decisions at milestone close — the milestone-close curation pass that promotes findings into the durable decision store. Triggers on "curate", "curate the milestone", "promote findings", "journal to codex", "promote to decisions", "curate reasoning log", "milestone curation", "promote hacks/risks/decisions". You are the driver Claude session: the scrum reconciler emits a `curation_proposed` event per closed-milestone task with candidate findings; you read the candidates, classify each as adr|glossary|pattern, human-gate each promotion with AskUserQuestion, and record it as a first-class scrum decision.