@juliusbrussee
Covers AI engineering, persona, and DevOps, known for caveman-commit, caveman-review, caveman-compress.
All Skills by @juliusbrussee
caveman
by @juliusbrussee
Caveman is a unique AI skill that enables your AI agents to communicate in a "caveman" style, drastically cutting down output tokens while maintaining 100% technical accuracy. By streamlining expressions, it effectively reduces LLM operational costs, boosts response speed by approximately 3x, and offers various compression levels, making it a powerful tool for optimizing AI interaction efficiency.
caveman-commit
by @juliusbrussee
Generates concise Git commit messages following Conventional Commits, emphasizing 'what' over 'why' and removing redundant phrasing.
caveman-review
by @juliusbrussee
Generates concise, actionable code review comments, one finding per line, including location, issue, and proposed fix.
caveman-compress
by @juliusbrussee
Compress natural language files into raw human language style, significantly reducing input tokens while retaining readable backups.
caveman-help
by @juliusbrussee
Displays reference cards in Caveman style, removing redundant words and filler, maintaining minimalist expression.
cavecrew
by @juliusbrussee
Cavecrew is a set of subagent presets designed to compress the output of Anthropic's default agents, significantly reducing context token consumption. It offers investigator, builder, and reviewer modes for code location, surgical edits, and bug checks, helping users avoid context exhaustion and improve development efficiency across multiple delegations.
caveman-stats
by @juliusbrussee
This skill provides immediate access to formatted statistics, allowing users to instantly view their AI usage. It requires no additional model processing, directly presenting key figures to help users quickly understand and analyze AI interaction patterns.
caveman-explore
by @juliusbrussee
This skill is a read-only repository explorer. When you face a cold-start exploration, need to locate a feature across many files, or a direct search has failed, it fans out parallel searches and reads, then returns compact path:line citations. It helps developers and AI agents find where code lives fast, saving tokens and time. It is designed for cold-start scenarios and for cases where a previous direct search has failed, avoiding wasted effort in unrelated files.
safe-refactor
by @juliusbrussee
Restructure code while preserving behavior. This skill defines behavior-preservation boundaries and establishes verification before structural edits, supporting extraction, consolidation, ownership moves, and cleanup. It keeps intermediate states buildable and testable, then runs the same proof after changes to ensure behavior matches and the requested structure is achieved, significantly reducing regressions and hidden bugs.
caveman-learn
by @juliusbrussee
This skill closes the loop on a caveman learn report by reviewing ranked token sinks and applying cost-lowering fixes—trimming heavy configuration files and offloading recurring context to a memory system—with per-edit consent. It never claims savings it hasn't measured and never makes the agent dumber. Use it when you run a learn report, want to lower your agent's token costs, need to trim a heavy prompt file, or want to stop re-pasting the same context every session. By keeping each accepted edit reversible and gated on actual measurement, it provides a safe, systematic approach to cutting token waste without sacrificing context quality.
caveman-optimize
by @juliusbrussee
This skill enables systematic evaluation of optimization observations, grounding decisions in evidence and operator choice, with paired evaluations to validate benefits before implementation. It helps developers avoid blind adoption, preventing regressions while ensuring real resource savings and quality preservation.
investigate-first
by @juliusbrussee
This skill helps developers diagnose ambiguous failures before editing code. It guides you to separate symptoms from root causes, trace inputs and state transitions, and rank hypotheses by evidence. By preventing untested edits, it cuts mean time to root-cause identification and reduces regression risk.
verify-and-stop
by @juliusbrussee
This skill translates acceptance conditions into the smallest sufficient proof set for validation-only tasks, completion checks, focused gate runs, and last-mile proof. It reuses still-current results, runs focused checks before wider gates, and distinguishes pass, fail, unavailable, and blocked precisely. It does not edit product code unless fixes are requested, and stops immediately when acceptance proof is complete, preventing scope creep and saving effort.
surgical-patch
by @juliusbrussee
This skill fixes bugs and small behavior changes at the narrowest responsible layer, preserving surrounding behavior and user modifications while avoiding unnecessary cleanup, renaming, or refactoring. Use it when regression proof and task-relevant tests matter. It helps developers quickly trace symptoms to root causes, minimize the change set, and verify fixes with focused tests, reducing regression risk and enhancing code stability. This leads to faster, safer bug fixes in real development.
caveman-discover
by @juliusbrussee
This skill automatically adds traceable labels to LLM workflows in a code repository (e.g., support reply, nightly digest), enabling cost monitoring per workflow. It inventories entry points, applies naming conventions, and integrates with the gateway through minimal changes, helping teams understand resource consumption of each AI task and optimize budget allocation.
lean-build
by @juliusbrussee
This skill is used for feature work with high overbuilding risk. It derives observable acceptance criteria and explicit non-goals from the request and repository, traces entry points through layers owning invariants, and reuses fitting seams to deliver a coherent end-to-end path. It helps developers maintain strict scope and an explicit stop condition, preventing over-engineering and improving delivery speed and code quality.
caveman-evidence-review
by @juliusbrussee
This skill performs read-only evidence review of Caveman platform data, basing conclusions on actual data rather than guesses. It separates measured costs, inferred headroom, and verified savings, and uses trace-level analysis to validate causality. Ideal for cost optimization and experiment evaluation workflows.
caveman-manage
by @juliusbrussee
This skill manages the lifecycle of eval-gated experiments, providing evidence-based approval recommendations and blocking unauthorized changes to ensure safe and compliant experiments. It is ideal for A/B testing, model experiments, and other scenarios where controlled releases are critical, helping teams reduce risk and improve management efficiency.
migration
by @juliusbrussee
This skill implements reversible and compatibility-safe transitions for schema, data, API, protocol, configuration, or dependency migrations. It ensures every change has explicit forward and rollback paths, preserves existing data, and makes destructive steps observable and separately authorized. It keeps mixed-version operation safe, following expand-migrate-verify-contract ordering. Use it to reduce migration risks and gain confidence in releasing changes.
caveman-setup
by @juliusbrussee
This skill guides developers in wiring their app's LLM callsites through a byte-preserving proxy gateway for usage and cost measurement. It provides verified integration patterns for OpenAI, Anthropic, and other SDKs, requiring only base URL and auth header changes—no refactoring. Ideal for teams wanting instant call monitoring, cost attribution, and secret safety across all live LLM endpoints.
compress
by @juliusbrussee
Compresses natural language files into Caveman style to reduce input Tokens, with the original file backed up as .original.md.
build
by @juliusbrussee
The "build" skill defines a single-threaded plan-execute process for an AI agent to implement specifications. It involves reading specs, parsing tasks, formulating detailed implementation plans (citing invariants, interfaces, files, and tests), then executing code modifications and running verifications. Upon failure, the agent triggers a backpropagation mechanism to analyze the root cause, deciding whether to fix the code or update the specification, ensuring each commit is verified and adheres to a strict policy.
spec
by @juliusbrussee
The `spec` skill is designed to be a robust technical specification management tool. It offers multiple modes to interact with specifications: generating new ones from user ideas, distilling them from existing code, or precisely amending current documentation. Crucially, this skill also features a backpropagation mechanism to automatically handle bugs and update affected specifications. By maintaining synchronization between specifications and the codebase, `spec` significantly improves software architecture clarity and overall engineering efficiency.
backprop
by @juliusbrussee
"backprop" is a structured bug-fixing process that ensures code quality and prevents recurrence by tracing, analyzing, updating specifications, and generating tests. It transforms each bug fix into an improvement of the system's specification, enhancing long-term software stability.
junior-to-senior
by @juliusbrussee
This Skill simulates an adversarial senior-engineer review of AI-generated plans, designs, and architectures. By conducting live codebase and web research, it diagnoses whether a plan is too vague or too granular, then rewrites it with evidence-backed, state-of-the-art practices. Use when you need to ensure plan quality and avoid outdated or hand-wavy output.
interface-kit
by @juliusbrussee
This Skill provides a systematic implementation guide for exceptional interfaces, covering a priority stack from accessibility, performance, typography, layout, color, motion to polish. It helps developers quickly deliver consistent, polished, and trustworthy UIs, eliminating endless CSS tweaks and reducing prototype-to-implementation time significantly.
context-canary
by @juliusbrussee
This skill plants a canary signal (e.g. user name and turn counter) at the beginning of every reply to monitor LLM context degradation in real time. When the signal fails, it automatically triggers alerts and a recovery protocol, helping users catch instruction drift, memory loss, and other silent failures early in long sessions.
fuck-slop
by @juliusbrussee
This skill detects and removes the telltale signs of AI-generated text (such as contrast structures, false ranges, regular cadence) and rewrites any text to sound like a specific person wrote it for a specific audience and genre. Ideal for creating natural, AI-free content for blogs, marketing copy, technical documentation, and more.
loop-factory
by @juliusbrussee
This Skill implements a spec-driven AI agent workflow that moves coding tasks as Markdown specs through inbox → active → archive states, with AI agents implementing and review gates ensuring quality. Use for repeatable, reviewable agent work.
caveman-es
by @juliusbrussee
Role-plays as a caveman, responding in minimalist Spanish, removing all technical terms, pleasantries, and filler words, delivering only core information.