Super AI Agent
Paste the following prompt into your AI chat to install this skill:
Please follow https://skillhub.cn/install/skillhub.md to install @user_0a48355b/weixiang.
About this skill
Problem
In ordinary Q&A, agents often flatten “deep thinking” into a single pass: either they repeat known context or fail on long research tasks. ThinkTank turns a topic into an evolvable reasoning domain, where the agent performs real inference and the engine only manages state. It fits brainstorming, multi-step reasoning, degraded long-task execution, and cross-session continuation, but not simple Q&A.
How it works
The skill runs an agent + engine loop: the agent thinks, the engine records and enforces thresholds.
- init sets maximum rounds and runtime config.
- Round 1 may use web search to build a baseline.
- Later rounds follow next-prompt, producing real next_state, novelty_score, and novelty_signals.
- Novelty is scored across information, perspective, pattern, and anomaly; scores at or above 0.6 with sufficient signals generate a structured discovery report.
- The engine sleeps when novelty dries up, direction converges, discoveries are sufficient, or the round cap is reached; state is persisted under .thinktank/ and resumed with wake.
Boundaries and notes
The agent must score each round honestly and must not fabricate next_state; web search is limited to the baseline round. Simple questions should not enter the engine. Long tasks should be split into multiple reasoning domains instead of forcing one session. If output is thin, increase the round cap or resume with wake.
Use Cases
- When a product idea is still vague, run several perspective-shifting reasoning rounds to surface non-consensus insights and report only high-novelty findings.
- For oversized research beyond one session, split the goal into domains, build a baseline, and converge each domain separately then summarize findings.
- When a literature review stalls in known summaries, establish a baseline, then evolve states to detect anomalous connections and record novelty signals.
- When a previous run stopped unexpectedly, wake the persisted ThinkTank state and continue the same reasoning domain without rerunning earlier rounds.
Best For
- Graduate researchers who need to turn a vague topic into testable, multi-round research hypotheses and anomaly-driven insights without repeating known literature.
- Product managers who must decompose long reporting tasks into resumable subdomains and synthesize domain reports across multiple sessions with persisted state.
- Domain researchers looking for anomalous connections beyond standard reviews and known baselines using novelty signals and structured reports each round.
- Prompt engineers who need to constrain agent loops, honest novelty scoring, and resumable state management without manual state edits or fabricated results.
Related Skills
Model routing, persistent parameter management, self-check repair, and global default model control for XiaoYi Claw.
Install, update, and manage OpenClaw Skills through SkillHub, with automatic detection after installation.
Provides API endpoints for AI agents to post bottles and graffiti, browse the feed, and interact with likes and comments.
ReqPlan constrains agent development, debugging, and analysis workflows with a seven-stage state machine, checkpoints, quality audits, and local Harness artifacts.