AI Engineering

Gryph Policy Authoring helps draft and test security policy rules that control what an AI coding agent may read, write, or run. Use it when you need to block, allow, warn, or require approval for agent actions.

Harness Step 3 Session Management creates init.sh, tasks.json, and progress.md to carry work state across sessions. Use it when an agent needs to resume tasks, track progress, and verify the environment quickly.

Heuresis Protocol widens an underexpanded idea field by generating diverse candidates in parallel without ranking or eliminating them. Use it when a topic is blank or has converged too early and you need more possibilities before choosing.

LangChain Tool Builder helps create LangChain Python tools with Claude Code’s fail-closed pattern. Use it when defining a new tool, adding validation or permissions, or setting up a ClaudeStyleTool base class.

Lite Constraints generates synthesis-constraints.json from per-run growth-factors.json and founder context. It is used after growth-factors mining and before synthesis to define white-space pairs, mandatory combinations, prohibited patterns, and category diversity rules.

Munger Mental Models applies Charlie Munger’s checklist, inversion, and bias checks to investments and decisions. Use it when evaluating a company, diagnosing misjudgment, or mapping useful mental models.

Onboard recommends an epistemic protocol from recent sessions, then can guide the user through scenario, trial, and quiz. Use it for new-user onboarding, re-onboarding, or protocol discovery.

Outsiders CEO Capital Allocation applies The Outsiders framework to judge management by per-share value creation, cash flow, and capital allocation choices. Use it for buybacks, acquisitions, dividends, leverage, and decentralization decisions.

Probe surfaces multiple candidate epistemic deficits for a vague situation and asks the user to recognize the best fit. Use it when something feels off but the right protocol is unclear.

QMT 内置因子回测根据策略描述、研报 PDF 或截图，解读因子与选股逻辑，并基于日频回测母版生成 QMT 策略编辑器内置回测脚本。适合截面因子选股、研报复现和 handlebar 调仓信号生成。

Realign fuses the project guide's current direction, external work signals, and the user's present understanding into one direction line. Use it when the guide no longer matches what the project is doing and a traced update is needed.

Recollect resolves vague recall by scanning contextual clues and presenting grounded excerpts for the user to recognize. Use it when you remember a discussion, topic, or concept but cannot name it yet.