Simply prompting an AI to "think like a senior" is often ambiguous, as it lacks clarity on elements such as problem definition, identifying potential failures, considering the impact of API changes, measuring performance, and verifying work before completion. The "Senior-thinking-skills" project addresses this by decomposing these judgment processes into 23 small, independent Agent Skills rather than a single, massive prompt. The overall structure consists of a router and disciplines, with `senior-engineer-mindset` serving as the router to select only the disciplines suitable for the current task.
The project includes the `clarify-the-real-problem` skill to identify the actual problem behind a request, and the `widen-the-solution-space` skill to ensure the AI isn't limited to a single initial implementation approach. It also features the `weigh-tradeoffs` skill to balance options against their level of reversibility, and the `premortem` skill to derive failure scenarios and pre-detection methods before implementation. Furthermore, it provides the `simplicity-budget` skill for managing the YAGNI principle and complexity budgets, the `interface-contracts` skill to review the promises made by exposed interfaces, and the `verifiability-first` skill for pre-defining success conditions and creating verifiable designs.
The `root-cause-discipline` skill ensures that causes and evidence are secured before modifying code, while the `surgical-change` skill minimizes changes unrelated to the request. The `measure-before-optimizing` skill mandates actual measurement before proceeding with performance optimization, the `verify-before-claiming` skill prohibits declaring completion without supporting evidence, and the `adversarial-review` skill focuses on finding errors rather than simply proving results. Separately, the method of utilizing sub-agents is isolated into the `delegate-to-subagents` skill, which handles details such as parallelization timing, the scale of context provided to workers, whether to grant write permissions to reviewers, recording the baseline SHA immediately before dispatch, and integrating outputs from multiple agents.
Skill Design Principles and Source Management
The core design principle is to combine multiple small skills of 30 to 70 lines, called upon as needed, rather than relying on a single, massive 1,100-line engineer prompt. The description section of each individual skill specifies when to use it, while the body describes only the specific procedures required for execution. The project was designed for traceability by analyzing various related projects such as Superpowers, agent-skills, ECC, and HNCS, with source information indicating the influence on each skill recorded in the README. Moreover, rather than simply imitating existing ideas, they were reconstructed to fit a concise discipline structure.
Functions that are impossible to implement through document-based skills alone were intentionally excluded. This was to avoid exaggerating "fake" features—such as an agent remembering things on its own or security enforcement being activated simply by installing a single document—where an actual runtime is absent. Areas requiring practical enforcement were isolated into a separate hook project. Consequently, the skills function to shift the agent's decision-making process, while the hooks serve to control limits when those judgments fail.
