@jiajiaoy

task-pilot provides a structured protocol for breaking complex multi-step work into verified subtasks with explicit checkpoints and context anchors, designed to prevent scope drift and survive context window compaction in long claude conversations.

thinkdeep enforces structured step-by-step reasoning (restate, explore, critique, select, confidence-check) to catch errors on complex questions before delivery. fits between problem clarification and task execution in agentic workflows.

learnloop captures user corrections, preferences, and project context into claude code's native auto-memory so subsequent sessions can reference learned patterns without re-teaching. memory entries are classified by type (user, feedback, project, reference) and indexed in a session-loaded manifest.

clarity-first surfaces hidden assumptions and real intent behind requests before execution, using a four-step protocol to score ambiguity and decide when to ask clarifying questions versus proceed.

honest-critic pushes claude to surface real flaws and reject bad ideas instead of defaulting to agreement. establishes a four-step protocol (red flag scan, steelman, prioritized pushback, verdict) with explicit anti-patterns to avoid.

signal-first calibrates response depth to question complexity, eliminating verbose preamble and filler. it provides a classification protocol and compression heuristics to match answer length to actual need.

lucky-today generates daily, weekly, monthly, and couple-mode fortune readings in chinese and english using deterministic offline rules. outputs lucky color, direction, number, item, dos/avoids, and scenario overlays for interviews, exams, travel. requires reading external rule files that are not provided in the skill document.

blind-box implements a gacha-style random collectible pull system across five themed series with rarity tiers and daily free pulls, but lacks substantive procedural detail, intent clarity, and failure mode handling.

story-english delivers serialized fiction across three genres (mystery, urban, sci-fi) at A2-B2 levels with embedded vocabulary and comprehension checks. narrative-driven learning without drills, but lacks interaction flow detail and failure recovery.

Use DeepSeek V4 (Flash & Pro) from the command line — one-shot Q&A, thinking mode, multi-turn chat. OpenAI-compatible API, no special CLI needed. Supports de...

Session coherence protocol for Claude — track decisions, resist scope creep, flag drift. Keeps long AI sessions on track without losing what was already deci...