Do not re-explain the same working method to an agent every session.
Skills Install Format License Last commit skills.sh
Install • Skills • How it works • State
This repository contains reusable SKILL.md workflows for coding assistants such as Codex, Claude Code, and other skill-compatible agents. Each skill is a small, versioned procedure with its own trigger description, constraints, and execution rules.
Quick start
Install one skill:
Or replace doc-style with another folder name from the list below.
Inspect what you are installing before using it:
Expected result: the selected skill is available to the compatible agent/runtime you install it into, and its name/description frontmatter tells that runtime when the workflow is relevant.
The problem
Prompts for recurring engineering work tend to drift: the same repository-cleanup rules, experiment discipline, documentation standard, or product-design process gets rewritten differently each time.
A skill turns that repeated instruction into a repository-owned procedure that can be reviewed, versioned, and reused.
Available skills
agent-grep — fast, token-thrifty structural code search, AST pattern matching, call graph tracing, and adaptive deduplication.
artifact — handling rules for PDFs, HTML reports, slides, LaTeX/XeLaTeX documents, and formula rendering.
chat-cleaner — map, name, and triage accumulated agent chat history into a clean inventory.
clean-system — consent-gated disk-space audit and cleanup workflow.
context-budgeting — token budget management, Context Ledger auditing, Ace extraction passes, and proactive context compaction.
doc-auditor — audit and prune AI self-talk, obsolete claims, and abandoned work logs from documentation.
doc-style — evidence-first README, specification, and launch-writing standard.
dumb-down — switch the agent to plain, non-technical explanation: no code, no jargon, no commands.
experiment-workflow — variable-isolated experiment workflow with explicit hypotheses and stress tests.
first-principles — engineering simplification workflow: question → delete → simplify → accelerate → automate.
gitlab — branch-protection and merge-request workflow guidance for GitLab repositories.
harness-eng — agent-project harness design: rules, progress state, feature tracking, and evaluator contracts.
lancedb-rag — semantic workspace indexing, dense vector search, BM25 full-text search, and hybrid recall with Reciprocal Rank Fusion (RRF).
latex-authoring — best practices for writing, compiling, and graph-indexing LaTeX academic papers.
multi-agent — multi-agent dispatch protocol: one task, one branch, one worktree, and coordinator integration.
pdf — visual-first workflow for reading, creating, and validating PDFs.
product-design — product architecture, user-flow design, stack selection, and frontend-first execution planning.
req-engineering — acceptance-criteria-first requirements discipline and verification trust hierarchy.
self-improving-loop — rules, architecture, and evaluation harness workflows for self-improving agent loops.
training-nn — neural-network training recipe and common failure-mode checklist.
vision — project orientation: locate or create the vision file that tells an agent what a project is before it edits anything.
word-farsi-docx — producing clean Word-native Persian (B Nazanin) .docx output without breaking formatting.
Personal machine configuration and third-party-authored skills are intentionally excluded from the public repository.
How it works
Each public skill lives in its own folder:
The file begins with YAML frontmatter:
The rest of the file defines the procedure, output contract, and guardrails. The repository is deliberately plain Markdown so every instruction is inspectable before it is installed.
Current state
Implemented and available
The skills listed above are present as standalone SKILL.md workflows and can be installed individually through the skills CLI.
Implemented but still under acceptance
Cross-runtime behavior is not yet continuously verified. A skill that parses correctly is not automatically proven to behave identically in every agent runtime.
The repository-owned public set is checked locally with:
This verifies the tracked skills' frontmatter and keeps the public lists in README.md and VISION.md aligned. It does not prove behavior across agent runtimes.
Planned
Compatibility checks across multiple supported agent runtimes.
More domain-specific engineering workflows when there is a repeated real use case.
Intentionally unsupported
Personal machine-specific configuration.
Private skills.
Third-party skills copied into this repository without a clear reason to redistribute them.
What sets this repository apart
These are design choices rather than novelty claims:
Procedures instead of prompt snippets. Each skill describes a repeatable workflow, not a one-off answer style.
One job per skill. A skill is scoped tightly enough to stay out of unrelated tasks.
Inspectable source. The behavior contract is plain Markdown in the repository.
Versionable working methods. Changes to how an agent should work can be reviewed like code.
Evals and test series
There is currently no repository-level CI proving cross-runtime compatibility. Run the local structural validator before committing changes to a public skill.
The smallest useful verification is:
install one skill into a supported runtime;
give the agent a task that clearly matches the skill description;
verify that the runtime discovers the skill;
check the agent's output against the procedure and guardrails in that SKILL.md.
Automating cross-runtime compatibility checks is planned work.
Future development
The next useful step is not adding many more skills. It is making the existing set easier to discover, install, and verify across runtimes.