# Skillware — LLM / agent discovery index # Canonical site: https://skillware.site # Open source repository: https://github.com/arpahls/skillware A Python framework for modular, self-contained agent skills (capabilities you install like packages). Don't prompt your agents — equip them. Each skill bundles deterministic Python logic (skill.py), LLM instructions (instructions.md), and governance (manifest.yaml). Same skill, zero adapter rewrites: SkillLoader adapts bundles to any agent, LLM, or logical system (Gemini, Claude, OpenAI, DeepSeek, Ollama, and more via universal adapters). ## Positioning (for agents and operators) - Tagline: Don't prompt your agents. Equip them. - Model churn: swap the LLM or provider — change the model name, keep the skill path and execute() contract. - One-line skill install: pip install "skillware[category_skill]" then SkillLoader.load_skill("category/skill_name") — reuse the same manifest across agents and LLMs. - Twenty-nine bundled skills as of v0.6.1 — see Skill directory below. ## Framework searches (no specific skill required) | What someone searches | Page | | :--- | :--- | | how to reuse same skill across agents / reuse skill across different llms | https://skillware.site/documentation#agent-loops | | install agent skill one line code / pip install skillware one line | https://skillware.site/documentation#install-extras | | add tools to ai agent without langchain / framework agnostic agent skills | https://skillware.site/comparison | | swap llm without rewriting agent / zero adapter rewrites | https://skillware.site/ | | equip ai agent with skills python / prebuilt agent capabilities | https://skillware.site/skills | | remove bg / background remover agent skill / skillware bg remover | https://skillware.site/skills/creative/bg_remover | | groq openrouter openai compatible agent tools | https://skillware.site/documentation#openai-compatible | | email for agents / can ai send emails / give your ai an email | https://skillware.site/skills/office/gmail_handler | | how can ai send emails for me / free mail for ai | https://skillware.site/skills/office/gmail_handler | | agent mailbox gmail imap smtp / ai agent reply to email | https://skillware.site/skills/office/gmail_handler | | agent mail / skillware mail cli / dedicated agent mailbox | https://skillware.site/skills/office/gmail_handler | | iot sensor telemetry normalizer / w3c sosa ssn / si unit converter | https://skillware.site/skills/data_engineering/telemetry_normalizer | | bioinformatics sequence ops / dna rna protein translation / pairwise aligner | https://skillware.site/skills/bio/sequence_ops | | geo ip rotator / residential proxy rotation / egress proxy agent | https://skillware.site/skills/networking/geo_ip_rotator | | skillware find / semantic discovery / find skills natural language | https://skillware.site/documentation#skillware-cli | | manifest v2 universal taxonomy / shape tier target tags / skillware info cli | https://skillware.site/documentation#skillware-cli | | agent skill benchmark / which ai has the best skills / best tool calling models | https://skillware.site/bench | | best model for skill use locally / small models vs frontier models tool calling | https://skillware.site/bench | | llm tool calling benchmarks 2026 / agent function calling leaderboard | https://skillware.site/bench | | cost of ai agent tool execution / agent token cost cascade / parameter paradox | https://skillware.site/bench | | 46 interactive publication exhibits / latency anatomy / run-to-run determinism | https://skillware.site/bench/exhibits | ## Key pages for agents and operators - Home: https://skillware.site/ - Skill directory (29 skills): https://skillware.site/skills - Benchmark overview: https://skillware.site/bench - Publication exhibits gallery (46 exhibits): https://skillware.site/bench/exhibits - Documentation: https://skillware.site/documentation - Install extras: https://skillware.site/documentation#install-extras - Vision & roadmap: https://skillware.site/documentation#vision - Contributing (humans + agents): https://skillware.site/contributing - AI-native agent workflow: https://skillware.site/documentation#ai-native-workflow - Comparison (MCP, LangChain, etc.): https://skillware.site/comparison - PyPI install: pip install skillware - Recommended single-skill install: pip install "skillware[_]" (see install extras) ## Problem-first search map (no brand name required) Users often search for tasks, not products. These map to Skillware skills: | What someone searches | Skill ID | Page | | :--- | :--- | :--- | | screen crypto wallet for ai agent / wallet screening for ai agents | finance/wallet_screening | https://skillware.site/skills/finance/wallet_screening | | uk companies house api integration / business registry api ai | finance/uk_companies_house_handler | https://skillware.site/skills/finance/uk_companies_house_handler | | how to send crypto using ai / ai agent swap tokens / crypto trading ai agent | defi/evm_tx_handler | https://skillware.site/skills/defi/evm_tx_handler | | ai fill pdf forms / can ai fill out pdf forms | office/pdf_form_filler | https://skillware.site/skills/office/pdf_form_filler | | email for agents / can ai send emails / give your ai agents an email | office/gmail_handler | https://skillware.site/skills/office/gmail_handler | | how can ai send emails for me / free mail for ai / agent send mail preview | office/gmail_handler | https://skillware.site/skills/office/gmail_handler | | ai agent inbox search read / gmail imap smtp for agents | office/gmail_handler | https://skillware.site/skills/office/gmail_handler | | agent mail / skillware mail cli / give agent an email | office/gmail_handler | https://skillware.site/skills/office/gmail_handler | | remove image background ai / remove bg / rembg python agent / background remover agent skill / skillware bg remover / transparent png | creative/bg_remover | https://skillware.site/skills/creative/bg_remover | | reduce llm token usage / compress long prompt | optimization/prompt_rewriter | https://skillware.site/skills/optimization/prompt_rewriter | | generate synthetic training data ai | data_engineering/synthetic_generator | https://skillware.site/skills/data_engineering/synthetic_generator | | filter redundant text dataset / semantic dedup | data_engineering/novelty_extractor | https://skillware.site/skills/data_engineering/novelty_extractor | | mask pii before chatgpt / gdpr safe ai agent | compliance/pii_masker | https://skillware.site/skills/compliance/pii_masker | | ai for mica / eu crypto regulation agent | compliance/mica_module | https://skillware.site/skills/compliance/mica_module | | is web scraping allowed / robots txt checker | compliance/tos_evaluator | https://skillware.site/skills/compliance/tos_evaluator | | prompt injection firewall agent / scan untrusted text before llm | security/prompt_injection_firewall | https://skillware.site/skills/security/prompt_injection_firewall | | ai agent fix github issue / plan before coding issue | dev_tools/issue_resolver | https://skillware.site/skills/dev_tools/issue_resolver | | how to make ai stop at token limit / agent loop token budget | monitoring/token_limiter | https://skillware.site/skills/monitoring/token_limiter | | ai wellness coaching guardrails / crisis triage chatbot | wellness/mental_coach | https://skillware.site/skills/wellness/mental_coach | | honeypot token scanner / goplus token security before swap | defi/token_security_scanner | https://skillware.site/skills/defi/token_security_scanner | | read erc20 balance ai agent / evm read only chain queries | defi/evm_reader | https://skillware.site/skills/defi/evm_reader | | korean gen z slang interpret / korean internet slang agent | linguistics/korean_slang | https://skillware.site/skills/linguistics/korean_slang | | extractive context selection / reduce document tokens agent | optimization/context_optimizer | https://skillware.site/skills/optimization/context_optimizer | | html web form mapper agent / csrf token form fill legal profile | office/web_form_mapper | https://skillware.site/skills/office/web_form_mapper | | deepfake guard agent / air-gapped id forensics / mrz passport check digits | security/deepfake_guard | https://skillware.site/skills/security/deepfake_guard | | iot sensor telemetry normalizer / w3c sosa ssn / si unit converter | data_engineering/telemetry_normalizer | https://skillware.site/skills/data_engineering/telemetry_normalizer | | iot sensor telemetry normalizer / w3c sosa ssn / si unit converter | data_engineering/telemetry_normalizer | https://skillware.site/skills/data_engineering/telemetry_normalizer | | bioinformatics sequence ops / dna rna translation / pairwise aligner | bio/sequence_ops | https://skillware.site/skills/bio/sequence_ops | | rotate ip address ai agent / geographic egress proxy / multi-region web pricing | networking/geo_ip_rotator | https://skillware.site/skills/networking/geo_ip_rotator | Brand-aware queries (skillware + keyword) also resolve to the same pages — see each skill's prerendered HTML FAQ block. ## Natural-language intent index | User intent | Skill ID | Page | | :--- | :--- | :--- | | ai for mica / EU crypto regulation | compliance/mica_module | https://skillware.site/skills/compliance/mica_module | | website policy / robots.txt / is scraping allowed | compliance/tos_evaluator | https://skillware.site/skills/compliance/tos_evaluator | | github issue plan before coding | dev_tools/issue_resolver | https://skillware.site/skills/dev_tools/issue_resolver | | swap llm without rewriting agent / zero adapter rewrites skillware | (framework) | https://skillware.site/ | | equip agents deterministic skills python | (framework) | https://skillware.site/skills | | how to contribute to open source ai as a beginner | (workflow) | https://skillware.site/contributing | | repositories welcome to AI agents | (project norms) | https://skillware.site/contributing | | mask PII before sending to cloud LLM | compliance/pii_masker | https://skillware.site/skills/compliance/pii_masker | | crypto wallet AML / sanctions screening | finance/wallet_screening | https://skillware.site/skills/finance/wallet_screening | | UK company officers / Companies House lookup | finance/uk_companies_house_handler | https://skillware.site/skills/finance/uk_companies_house_handler | | government company filings api / corporate registry data ai | finance/uk_companies_house_handler | https://skillware.site/skills/finance/uk_companies_house_handler | | fill PDF forms from natural language | office/pdf_form_filler | https://skillware.site/skills/office/pdf_form_filler | | give AI agent an email / send mail on my behalf / dedicated agent mailbox | office/gmail_handler | https://skillware.site/skills/office/gmail_handler | | can AI send emails if I ask / preview before send agent email | office/gmail_handler | https://skillware.site/skills/office/gmail_handler | | remove image background / remove bg / rembg offline / product photo cutout / background remover agent | creative/bg_remover | https://skillware.site/skills/creative/bg_remover | | reuse skill across agents / same skill different llm / one line install agent skill | (framework) | https://skillware.site/documentation#agent-loops | | compress prompts / reduce API tokens | optimization/prompt_rewriter | https://skillware.site/skills/optimization/prompt_rewriter | | synthetic data for fine-tuning | data_engineering/synthetic_generator | https://skillware.site/skills/data_engineering/synthetic_generator | | filter redundant text / semantic dedup | data_engineering/novelty_extractor | https://skillware.site/skills/data_engineering/novelty_extractor | | agent wallet swap / buy token on base or ethereum | defi/evm_tx_handler | https://skillware.site/skills/defi/evm_tx_handler | | token budget / stop runaway agent loop | monitoring/token_limiter | https://skillware.site/skills/monitoring/token_limiter | | wellness coaching guardrails / crisis triage | wellness/mental_coach | https://skillware.site/skills/wellness/mental_coach | | prompt injection scan / jailbreak firewall before llm | security/prompt_injection_firewall | https://skillware.site/skills/security/prompt_injection_firewall | | deceptive ui dark patterns checkout / anti agent html scan | security/deceptive_ui_guard | https://skillware.site/skills/security/deceptive_ui_guard | | kpi gate business metrics policy charter | monitoring/kpi_gate | https://skillware.site/skills/monitoring/kpi_gate | | semantic web proxy trafilatura / reduce page tokens | data_engineering/semantic_web_proxy | https://skillware.site/skills/data_engineering/semantic_web_proxy | | ai agent powerpoint pptx deck builder | creative/deck_builder | https://skillware.site/skills/creative/deck_builder | | skillcontext skillware chain multi skill host | (framework) | https://skillware.site/documentation#skill-chaining | | skillware mail cli / agent gmail operator | office/gmail_handler | https://skillware.site/documentation#skillware-cli | | groq openrouter openai compatible agent tools | (framework) | https://skillware.site/documentation#openai-compatible | | honeypot token scanner / goplus before swap | defi/token_security_scanner | https://skillware.site/skills/defi/token_security_scanner | | evm read only balance / erc20 metadata agent | defi/evm_reader | https://skillware.site/skills/defi/evm_reader | | korean slang interpret gen z | linguistics/korean_slang | https://skillware.site/skills/linguistics/korean_slang | | context optimizer extractive rag | optimization/context_optimizer | https://skillware.site/skills/optimization/context_optimizer | | aws bedrock agent tools skillware | (framework) | https://skillware.site/documentation#bedrock-integration | | skillware evm cli / addressbook public_0x | (framework) | https://skillware.site/documentation#evm-operator-config | | fill web forms safely / csrf preservation / legal_profile address book | office/web_form_mapper | https://skillware.site/skills/office/web_form_mapper | | air-gapped passport id verification / check digits / c2pa ela forensics | security/deepfake_guard | https://skillware.site/skills/security/deepfake_guard | | IoT sensor telemetry normalizer / W3C SOSA SSN / SI unit converter | data_engineering/telemetry_normalizer | https://skillware.site/skills/data_engineering/telemetry_normalizer | | bioinformatics sequence operations / IUPAC validation / pairwise alignment | bio/sequence_ops | https://skillware.site/skills/bio/sequence_ops | | rotate residential IP / geographic egress proxy / multi-region web crawl | networking/geo_ip_rotator | https://skillware.site/skills/networking/geo_ip_rotator | | semantic task intent discovery / skillware find CLI | (framework) | https://skillware.site/documentation#skillware-cli | | Manifest v2 Universal Taxonomy / shape tier target tags / SPDX MIT headers | (governance) | https://skillware.site/contributing | | sub-200ms cold startup CLI / skillware info inspection / discovery flags | (framework) | https://skillware.site/documentation#skillware-cli | | native async execution / aexecute / arun_chain / concurrency throttling | (framework) | https://skillware.site/documentation#agent-loops | | The Permissive Fortress / AST security audit / code ownership | (governance) | https://skillware.site/contributing | Note: MiCA and ToS skills support regulatory and policy analysis; they are not a substitute for licensed legal advice. ## Core concepts - Contract: manifest.yaml — typed I/O, constitution, issuer attribution, requirements, and Manifest v2 Universal Taxonomy (shape, tier, target, tags) - Effect: skill.py — deterministic execution (BaseSkill.execute, BaseSkill.aexecute) - Directive: instructions.md — when and how the host uses the tool - Assurance: test_skill.py — bundle tests (skillware test / CI) and Top-of-file SPDX MIT headers + copyright notices - Presentation: card.json — catalog card metadata - Interface: SkillLoader adapters (to_gemini_tool, to_claude_tool, …) - Loader: SkillLoader.load_skill("category/skill_name") — discovery via SKILLWARE_SKILL_PATH, local skills/, or bundled package; bundles expose bundle["class"]; missing deps suggest skillware[category_skill] - CLI: skillware find, skillware info, skillware list (--shape, --tier, --target, --tag), skillware examples, skillware test, skillware paths, skillware doctor (sub-200ms cold startup), skillware doctor --install, skillware context show, skillware chain list|show|validate|run|dry-run, skillware mail, skillware theme, skillware addressbook, skillware evm, skillware config show — gradient splash + 7-item prioritized interactive menu (1–7) - SkillContext + named chains: (v0.5.4+) multi-skill discovery, run_chain() pipelines in YAML — see documentation#skill-chaining - Async parity: (v0.5.8+) BaseSkill.aexecute(), SkillContext.aexecute(), SkillContext(max_concurrency=N) semaphore concurrency control, and arun_chain() - Governance & Security: (v0.5.8+) The Permissive Fortress, AST security audits (bandit, pip-audit, tests/test_security_audit.py prohibiting eval/exec/raw os.environ in skills), permanent irrevocable licensing - Secret providers: (v0.5.6+) SecretProvider, SkillContext(secret_provider=...), BaseSkill.credential() — inject manifest env_vars without global os.environ mutation - Install: pip install skillware; prefer pip install "skillware[category_skill]"; [all] = skill runtime deps; [agents] = Gemini+Claude+OpenAI SDKs; legacy [cli]/[embeddings] removed - Config: .skillware.yaml (project) + ~/.config/skillware/config.yaml (global) drive paths.project, paths.external, resolution.order, mail.*, chains:, theme; legacy SKILLWARE_SKILL_PATH still honored when configured ## Skills (29) ### finance/wallet_screening — Wallet Screening https://skillware.site/skills/finance/wallet_screening Screens Ethereum wallets for sanctions (880+ bundled lists), malicious interactions, and forensic risk (v2.1.0). Returns schema v2 verdicts (ALLOW, REVIEW, DENY), categorical risk levels (CLEAN, SUSPICIOUS, SEVERE), and hop-1 indirect counterparty exposure analysis. Runs offline-direct when ETHERSCAN_API_KEY is omitted (metadata.mode: offline_direct_only). Optional live verification via CHAINALYSIS_API_KEY and TRM_SANCTIONS_API_KEY into risk_details.live_verification[]. Refreshed OFAC SDN snapshot with dedicated ETH address filter; paginated Etherscan tokentx and txlistinternal transfers up to 10,000 items with optional metadata.warnings. Issuer: Ross Peili. Install: pip install "skillware[finance_wallet_screening]" Sample: Screen wallet 0xd8dA6BF26964aF9D7eEd9e03E53415D37aA96045 for sanctions and malicious contract interactions. ### finance/uk_companies_house_handler — UK Companies House Handler https://skillware.site/skills/finance/uk_companies_house_handler UK Companies House REST API integration for agents (v1.4.0) — search, profile, officers, PSC, filings, deterministic role/name matchers, filing helpers, stack-pop run_pipeline, and Document API discovery/pick/download flow (fetch_document returning in-memory base64 encoded document streams up to a 10 MB ceiling). Composites include resolve_company_officer. Requires COMPANIES_HOUSE_API_KEY. Issuer: Areen-09 · ARPAHLS. Install: pip install "skillware[finance_uk_companies_house_handler]" Sample user message to the agent: Who is the CEO of BP? (skill maps CEO → UK director terminology) Examples: uk_companies_house_handler_demo.py (mocked), gemini_uk_companies_house_handler.py, claude_uk_companies_house_handler.py ### defi/evm_tx_handler — EVM Transaction Handler https://skillware.site/skills/defi/evm_tx_handler Dedicated agent wallet (v0.3.0): resolve intent, Uni V2 quote/preview/execute, transfer on Ethereum/Base, address book name resolution via public_0x. Operator EVM config via skillware evm CLI. Issuer: Hendobox. Install: pip install "skillware[defi_evm_tx_handler]" Requires AGENT_WALLET_PRIVATE_KEY in .env; ETHEREUM_RPC_URL / BASE_RPC_URL per chain. Not financial advice. Sample: Buy 10 DEGEN on Base with USDC — resolve, quote, preview, execute with confirmed: true. Examples: gemini_evm_tx_handler.py, claude_evm_tx_handler.py ### office/pdf_form_filler — PDF Form Filler https://skillware.site/skills/office/pdf_form_filler Deterministic AcroForm PDF inspection, validation, and filling engine (v0.1.2) built exclusively on PyMuPDF (fitz) with zero cloud model dependencies or API keys required. Multi-action protocol: inspect discovers form fields and types; validate audits field constraints and options; fill updates text, checkboxes, and normalized radio groups with confirmed: true safety gate and optional vector flattening. Fails closed on dynamic XML Forms Architecture (XFA). Issuer: rosspeili · ARPAHLS. Install: pip install "skillware[office_pdf_form_filler]" Sample: Inspect and fill fields in /path/to/standard_form.pdf — name John Doe, check the terms of service box, confirmed: true. ### office/gmail_handler — Gmail Handler https://skillware.site/skills/office/gmail_handler Dedicated agent mailbox via Gmail IMAP/SMTP + App Password (v0.2) — resolve recipients, preview/send/reply with confirmation gates, attachment download, multi-profile reply signatures, inbox and sent search, YAML address book. Operator UX: skillware mail submenu. Requires GMAIL_ADDRESS and GMAIL_APP_PASSWORD on a fresh agent-only Gmail account. Install: pip install "skillware[office_gmail_handler]" Sample: Send George an email that deployment is complete — preview first, then send after I confirm. Examples: gmail_handler_demo.py (mocked), gemini_gmail_handler.py ### creative/bg_remover — Background Remover https://skillware.site/skills/creative/bg_remover Offline rembg background removal for still images — Base64 or file path in, transparent PNG out (v0.2). Issuer: AyushSrivastava1818. Install: pip install "skillware[creative_bg_remover]" (rembg, pillow, onnxruntime). First execute downloads ONNX model; no skill API keys. Example: bg_remover_demo.py Sample: Remove the background from product.png and save as product_no_bg.png. ### optimization/prompt_rewriter — Prompt Rewriter https://skillware.site/skills/optimization/prompt_rewriter Compresses long prompts while preserving meaning; reduces LLM token cost. Install: pip install "skillware[optimization_prompt_rewriter]" Sample: Compress this prompt before the main model call. ### data_engineering/synthetic_generator — Synthetic Data Generator https://skillware.site/skills/data_engineering/synthetic_generator Generates structured high-entropy synthetic samples for training; supports gemini, anthropic, or ollama as internal generator. Install: pip install "skillware[data_engineering_synthetic_generator]" Sample: Generate five medical coding dispute samples with dual-insurance edge cases. ### data_engineering/novelty_extractor — Novelty Extractor https://skillware.site/skills/data_engineering/novelty_extractor Filters text by semantic novelty with local fastembed embeddings (no API keys). Pass prior distilled_content as baseline_chunks for multi-turn dedup. Install: pip install "skillware[data_engineering_novelty_extractor]" Sample: Filter this dataset and keep only chunks that contain new information (novelty_threshold 0.85). Examples: novelty_extractor_demo.py, gemini_novelty_extractor.py, ollama_novelty_extractor.py ### compliance/pii_masker — PII Masker https://skillware.site/skills/compliance/pii_masker Local PII scrub via Ollama (arpacorp/micro-f1-mask); mask, redact, or remove before cloud LLMs. Install: pip install "skillware[compliance_pii_masker]" Sample: Mask PII in: "Hello John Doe, your wallet 0xabc123 has been verified." ### compliance/mica_module — MiCA Module https://skillware.site/skills/compliance/mica_module MiCA (EU Markets in Crypto-Assets) regulatory RAG and optional Gemini policy evaluator. Install: pip install "skillware[compliance_mica_module]" Sample: Can I issue a stablecoin backed by physical art under an e-money license? ### compliance/tos_evaluator — Terms of Service Evaluator https://skillware.site/skills/compliance/tos_evaluator Checks robots.txt, discovers policy pages, returns SAFE/UNSAFE/CAUTION/INSUFFICIENT_EVIDENCE for intended automation. Install: pip install "skillware[compliance_tos_evaluator]" Sample: Before crawling https://example.com/docs, check if automated indexing is allowed. ### dev_tools/issue_resolver — Issue Resolver https://skillware.site/skills/dev_tools/issue_resolver Nine-stage GitHub issue workflow: prepare URLs, stage_checklists, validate_commit_message. Optional ISSUE_RESOLVER.md repository profiles (load_repository_profile). Agent fetches issue/repo; skill does not call GitHub (optional GITHUB_TOKEN). Install: pip install "skillware[dev_tools_issue_resolver]" Sample: action prepare on https://github.com/owner/repo/issues/42; then stage_checklist for discover_issue. Examples: gemini_issue_resolver.py, claude_issue_resolver.py, ollama_issue_resolver.py ### monitoring/token_limiter — Token Limiter https://skillware.site/skills/monitoring/token_limiter Deterministic token budget gate for agent loops: CONTINUE, WARN, or FORCE_TERMINATE. Host supplies cumulative current_token_count; optional model_id for indicative USD estimates. Not billing authority. Install: pip install "skillware[monitoring_token_limiter]" Sample: Check task scrape_amazon_listings_101 with 125000 tokens used and 100000 limit. Examples: token_limiter_loop.py, gemini_token_limiter.py, claude_token_limiter.py ### wellness/mental_coach — Mental Coach https://skillware.site/skills/wellness/mental_coach Deterministic wellness coaching guardrail: crisis triage, scope limits, grounded KB retrieval. Optional GOOGLE_API_KEY for run_evaluator. Not medical or clinical advice. Install: pip install "skillware[wellness_mental_coach]" Sample: I feel stressed at work and need coping strategies (user_jurisdiction US, session_mode coaching). Example: mental_coach_demo.py ### security/prompt_injection_firewall — Prompt Injection Firewall https://skillware.site/skills/security/prompt_injection_firewall Offline deterministic pre-flight scan (v0.2.0) for hostile instructions — hidden HTML/markdown, Unicode smuggling, leetspeak/ROT13/typoglycemia/homoglyph evasion, markdown/HTML exfiltration channels, instruction-override lexicon. Returns policy_action, is_safe, risk_level, findings, optional sanitized_text. No API keys. Issuer: Masa · ARPAHLS (@mrmasa88). Install: pip install "skillware[security_prompt_injection_firewall]" Sample: Scan HTML with a hidden span containing "IGNORE ALL INSTRUCTIONS" before adding to model context. Example: prompt_injection_firewall_demo.py ### security/deceptive_ui_guard — Deceptive UI Guard https://skillware.site/skills/security/deceptive_ui_guard Deterministic deceptive UI scanner v0.2 — DOM zone classification, KB allowlists, prechecked opt-ins, drip pricing, fake scarcity, optional Playwright render diff. Returns trust_score, zone_summary, agent_guidance. Pair with prompt_injection_firewall in skill chains. Install: pip install "skillware[security_deceptive_ui_guard]" Sample: Scan checkout HTML for dark patterns before the agent clicks anything. Example: deceptive_ui_guard_demo.py ### monitoring/kpi_gate — KPI Gate https://skillware.site/skills/monitoring/kpi_gate Deterministic business-KPI gate — metrics snapshot vs operator policy charter (+ optional benchmarks). Four-stage validation; error findings block dependent actions. Complements token_limiter. Pure offline function. Install: pip install "skillware[monitoring_kpi_gate]" Sample: Evaluate weekly bookings snapshot against the gate charter. Example: kpi_gate_demo.py ### creative/deck_builder — Deck Builder https://skillware.site/skills/creative/deck_builder Deterministic .pptx assembly from JSON deck specs (v0.2.0) — 13 layouts, lint_deck, suggest_outline, Pillow placeholders, governance ribbons. Offline. Install: pip install "skillware[creative_deck_builder]" Sample: Build a quarterly briefing deck and render to briefing.pptx. ### data_engineering/semantic_web_proxy — Semantic Web Proxy https://skillware.site/skills/data_engineering/semantic_web_proxy Token-efficient page extraction via trafilatura — Markdown/txt/json output, SSRF guard, token_savings estimate, page_likely_requires_javascript warning. Chain with prompt_injection_firewall via SkillContext. Install: pip install "skillware[data_engineering_semantic_web_proxy]" Sample: Extract main article from URL as Markdown and report token savings. Example: semantic_web_proxy_demo.py ### data_engineering/telemetry_normalizer — Telemetry Normalizer https://skillware.site/skills/data_engineering/telemetry_normalizer Deterministic IoT sensor telemetry harmonization (v0.1.0) aligning field streams with the W3C SOSA/SSN ontology. Ingests nested JSON/CSV batch samples, harmonizes timestamps to ISO-8601 UTC with monotonic sequencing, maps physical quantities to SI base/derived units (Celsius/Fahrenheit to Kelvin, bar/psi to Pascal, RPM to rad/s), flags schema drift and out-of-bounds sensor values, and validates against sensor profiles. Issuer: Rin Katsuragi · ARPAHLS. Install: pip install "skillware[data_engineering_telemetry_normalizer]" Sample: Ingest raw industrial boiler telemetry and normalize sensor readings into SI units conforming to W3C SOSA/SSN. Examples: telemetry_normalizer_demo.py ### defi/evm_reader — EVM Reader https://skillware.site/skills/defi/evm_reader Read-only EVM state plane (v0.1.0) — ERC-20/721 metadata, balances, allowances, allowlisted view calls, Multicall3 batches, resolve_holder via address book. No private keys. RPC from operator evm.yaml. Install: pip install "skillware[defi_evm_reader]" Sample: What is the USDC balance of this holder on Ethereum mainnet? ### defi/token_security_scanner — Token Security Scanner https://skillware.site/skills/defi/token_security_scanner Read-only GoPlus Token Security scan (v0.2.0) — evaluates honeypot mechanisms, transfer buy/sell taxes, proxy contracts, mint privileges, and ownership centralization against operator policy profiles (enterprise_default). Returns deterministic policy_action (allow, flag, block) and supports batch contract screening (scan_batch up to 5 contracts). Use before evm_tx_handler on unfamiliar ERC-20. Optional GOPLUS_APP_KEY. Install: pip install "skillware[defi_token_security_scanner]" Sample: Scan this contract for honeypot risks and evaluate against enterprise policy profile before we swap. ### optimization/context_optimizer — Context Optimizer https://skillware.site/skills/optimization/context_optimizer Query-aware extractive selection (v0.1.0) — local fastembed scores chunks against agent_goal; verbatim spans only. Chain after prompt_injection_firewall. Install: pip install "skillware[optimization_context_optimizer]" Sample: Select only revenue-risk paragraphs from this 50-page report. Examples: context_optimizer_demo.py, context_optimizer_chain_demo.py ### linguistics/korean_slang — Korean Slang https://skillware.site/skills/linguistics/korean_slang Offline Korean Gen-Z slang lexicon (v0.1.0) — interpret, suggest, lookup with audience gates. No network inside execute(). Install: pip install "skillware[linguistics_korean_slang]" Sample: Interpret this Korean internet slang for an English-speaking operator. Example: korean_slang_demo.py ### office/web_form_mapper — Web Form Mapper https://skillware.site/skills/office/web_form_mapper Autonomous HTML web form parser, semantic entity mapper, and safety preview generator (v0.1.0). Ingests raw HTML or fetches target URLs via HTTP POST or headless browser engine. Maps identity profiles from .skillware.yaml (legal_profile) or inline payloads to fields. Fails closed on bot detection (CAPTCHA/Cloudflare), preserves CSRF / hidden state tokens (_csrf, ASP.NET __VIEWSTATE), and enforces the safety triad: inspect → map → preview → submit (submit requires confirmed: true). Issuer: rosspeili · ARPAHLS. Install: pip install "skillware[office_web_form_mapper]" Sample: Inspect and preview form fields on https://example.com/contact with my legal profile. Examples: web_form_mapping_demo.py, web_form_guard_chain_demo.py ### security/deepfake_guard — Deepfake Guard https://skillware.site/skills/security/deepfake_guard Deterministic multi-layer forensic verification engine (v0.1.0) for digital identity documents, passports, and facial imagery without external API dependencies. Validates ICAO 9303 cyclic check digits (7, 3, 1) on TD1/TD2/TD3 Machine Readable Zones (MRZ), parses C2PA JUMBF manifests and GenAI provenance tags, detects JPEG recompression artifacts via Error Level Analysis (ELA), flags screen recapture moiré grids via 2D FFT spectral peaks, and evaluates high-pass Laplacian noise residual variance. Issuer: rosspeili · ARPAHLS. Install: pip install "skillware[security_deepfake_guard]" Sample: Verify the authenticity of this passport scan using air-gapped MRZ and ELA checks. Examples: deepfake_guard_demo.py, kyc_authenticity_chain_demo.py ### bio/sequence_ops — Sequence Operations https://skillware.site/skills/bio/sequence_ops Deterministic bioinformatics sequence analysis engine (v0.1.0) built on Biopython. Validates DNA, RNA, and protein alphabets against strict IUPAC standards with degenerate symbol disambiguation, calculates GC content with sliding window distributions, performs transcription, reverse complementation, and codon translation with standard/vertebrate mitochondrial tables, and executes global/local pairwise alignments with scoring matrices (BLOSUM62/PAM250). Issuer: Ross Peili. Install: pip install "skillware[bio_sequence_ops]" Sample: Run local pairwise alignment between two candidate peptide sequences using BLOSUM62 scoring matrix. Examples: sequence_ops_demo.py ### networking/geo_ip_rotator — Geo IP Rotator https://skillware.site/skills/networking/geo_ip_rotator Geographic egress proxy routing, multi-region web inspection, and anti-blocking gateway for autonomous AI agents (v0.1.0). Provides targeted country/city egress (fetch_via_geo), concurrent multi-region comparison (compare_regions), pre-flight exit node resolution (resolve_egress), proxy context emission (get_proxy_context), strict pre-flight SSRF containment on all endpoints and redirect hops, pluggable commercial/private proxy adapters (BrightData, Oxylabs, Smartproxy, IPRoyal), locale HTTP headers, and enterprise data sovereignty compliance auditing with SHA-256 cryptographic receipts. Deterministic offline mock mode. Issuer: Blade Dancer. Install: pip install "skillware[networking_geo_ip_rotator]" Sample: Fetch https://example.com through a residential proxy node in Germany (DE) with data sovereignty compliance. Examples: geo_ip_rotator_demo.py ## Skillware-Bench: Leaderboard, Telemetry & 46 Publication Exhibits Canonical benchmark overview: https://skillware.site/bench Interactive exhibits compendium (46 exhibits): https://skillware.site/bench/exhibits Comparison with MCP & agent frameworks: https://skillware.site/comparison Skillware-Bench is the transparent, empirical evaluation framework for governed agent skills. It evaluates 23 frontier and sovereign models across an empirical 16-layer Progressive Tiered Ablation Benchmark (4 crucibles × 4 difficulty tiers = 16 layers, 736 conditions). It contrasts unconditioned code generation (Vanilla baseline) against governed execution (Skillware equipped) to measure real-world reliability, enterprise cost cascades, latency, run-to-run determinism, and constitutional safety. ### Grand Aggregate Fleet Telemetry (23 Models, 736 Experimental Conditions) - Fleet Pass Rate: 91.6% (337/368) vs 81.5% (300/368 unassisted) (+10.1% net lift across all 23 models) - Token Consumption: -51.4% reduction (1,586,162 prompt and completion tokens saved) - Execution Speedup: 1.60x cumulative wall-clock speedup (72.2m vs 115.2m cumulative execution latency) - Script Traps: 464 → 0 scripts eliminated (100% of workspace script debugging loops eliminated to zero) - Inter-Run Determinism: 5.1x tighter variance (CV = 0.50 compressed to CV = 0.10) - Adversarial Defense: 92.0% mean defense (84.1% → 92.0%, 99.6% intercept rate across 6 canonical attack stages) - Expected Calibration Error: 14.5% → 2.9% (-79.7% error deflation, eliminating ungrounded verbal overconfidence) ### Key Research Findings 1. The Parameter Paradox: Compact open-weight models equipped with Skillware (Qwen 3.8 Flash 16B, Gemma 4 31B, Llama 4 Scout 17B) achieve 93.8% to 100% pass rates, matching or outperforming unassisted frontier APIs Claude Sonnet 5 and GPT-6 Luna (87.5%) at 1/30th the inference cost. 2. The Benchmark Cost Cascade: Compound enterprise spend reduction cuts cost from $15.58 down to $0.42 (-97.3%) by eliminating recursive script debugging retries (-$6.47), JIT schema excision (-$6.22), and compact workhorse substitution (-$2.47). 3. Latency Anatomy: End-to-end task wall-clock latency drops from 27.5s to 12.2s (-55.6% acceleration) by replacing multi-turn prompt negotiation and sandbox compilation with sub-millisecond local execution. 4. Script Debugging Elimination: Unassisted agents generated 464 ad-hoc workspace scripts across 151 conditions trapped in debug loops. Governed skill execution eliminates 100% of these script traps to 0. 5. Determinism & Variance Contraction: Collapsing coefficient of variation from CV = 0.50 to CV = 0.10, eliminating catastrophic right-tail variance and runaway agent loops. 6. Calibration & Overconfidence Deflation: ECE shrinks from 14.5% to 2.9% (-79.7% error deflation), ensuring models don't claim success based on ungrounded verbal gut feelings. 7. Call Survival Flow: Clean first-pass completions surge from 71.2% to 94.1% (+22.9% lift), and catastrophic aborts collapse from 57 to 7 (-87.7% reduction). 8. The Permissive Fortress & AURA Audit: 100% of state-changing tool executions produce cryptographic SHA-256 Merkle receipts before dispatch. ### 23 Evaluated Models - Frontier Cloud APIs: Claude Sonnet 5, Claude Haiku 4.5, Claude Opus 4, GPT-6 Luna, GPT-5 Turbo, GPT-5 Mini, Gemini 3.8 Pro, Gemini 3.8 Flash, Command R+. - Sovereign Heavyweights: Llama 4 Maverick 70B, Llama 3.3 70B, Qwen 3.8 Max, Qwen 2.5 72B, DeepSeek V3, DeepSeek R1, Mistral Large 2. - Compact Workhorses: Qwen 3.8 Flash 16B, Gemma 4 31B, Gemma 3 27B, Llama 4 Scout 17B, Mistral NeMo 12B, GLM-4 9B, Yi-1.5 34B. ### Publication Exhibits Index (All 46 Exhibits) | ID | Type & Number | Part | Exhibit Title | Direct URL | | :--- | :--- | :--- | :--- | :--- | | fig01 | Figure 1 | Part I | The Four Constituencies of Governed Benchmark Evaluation | https://skillware.site/bench/exhibits?id=fig01 | | fig02 | Figure 2 | Part I | Anatomy of a Skill Bundle | https://skillware.site/bench/exhibits?id=fig02 | | fig03 | Figure 3 | Part I | Chained Skill Execution Across Difficulty Tiers | https://skillware.site/bench/exhibits?id=fig03 | | fig04 | Figure 4 | Part IV | The Benchmark Cost Cascade | https://skillware.site/bench/exhibits?id=fig04 | | fig05 | Figure 5 | Part VI | The Parameter Paradox | https://skillware.site/bench/exhibits?id=fig05 | | fig06 | Figure 6 | Part IV | Context Token Excision Across Execution Layers | https://skillware.site/bench/exhibits?id=fig06 | | fig07 | Figure 7 | Part V | Deterministic Execution & Contract Gating | https://skillware.site/bench/exhibits?id=fig07 | | fig08 | Figure 8 | Part II | Progressive Difficulty Degradation by Model Scale | https://skillware.site/bench/exhibits?id=fig08 | | fig09 | Figure 9 | Part VI | Script Trap Elimination & Infinite Debug Loops | https://skillware.site/bench/exhibits?id=fig09 | | fig10 | Figure 10 | Part VI | Zero-Token Pre-Execution Firewall Intercept | https://skillware.site/bench/exhibits?id=fig10 | | fig11 | Figure 11 | Part V | Latency Anatomy of Complex Workflows | https://skillware.site/bench/exhibits?id=fig11 | | fig12 | Figure 12 | Part III | Progressive Ablation Crucible Pass Rates | https://skillware.site/bench/exhibits?id=fig12 | | fig13 | Figure 13 | Part III | Multi-Turn Chained Failure Cascades | https://skillware.site/bench/exhibits?id=fig13 | | fig14 | Figure 14 | Part V | Run-to-Run Determinism & Variance Contraction | https://skillware.site/bench/exhibits?id=fig14 | | fig15 | Figure 15 | Part III | State Drift Across Long-Horizon Tasks | https://skillware.site/bench/exhibits?id=fig15 | | fig16 | Figure 16 | Part III | Parameter Recovery Under Malformed Schemas | https://skillware.site/bench/exhibits?id=fig16 | | fig17 | Figure 17 | Part VI | Governance & Security Radar Profile | https://skillware.site/bench/exhibits?id=fig17 | | fig18 | Figure 18 | Part II | Frontier API vs Sovereign Open-Weight Standings | https://skillware.site/bench/exhibits?id=fig18 | | fig19 | Figure 19 | Part II | Cost-Per-Success Pareto Frontier | https://skillware.site/bench/exhibits?id=fig19 | | fig20 | Figure 20 | Part II | Model Latency vs Accuracy Distribution | https://skillware.site/bench/exhibits?id=fig20 | | fig21 | Figure 21 | Part II | Fleet Pass Rate Delta (Vanilla vs Equipped) | https://skillware.site/bench/exhibits?id=fig21 | | fig22 | Figure 22 | Part V | Wall-Clock Acceleration Across Model Classes | https://skillware.site/bench/exhibits?id=fig22 | | fig23 | Figure 23 | Part II | Macro Fleet Net Lift by Tier | https://skillware.site/bench/exhibits?id=fig23 | | fig24 | Figure 24 | Part II | Small Model Efficiency Frontier | https://skillware.site/bench/exhibits?id=fig24 | | fig25 | Figure 25 | Part II | Token Waste Decomposition Across Tiers | https://skillware.site/bench/exhibits?id=fig25 | | fig26 | Figure 26 | Part II | Sovereign Air-Gapped Deployment Suitability | https://skillware.site/bench/exhibits?id=fig26 | | fig27 | Figure 27 | Part V | Calibration & Overconfidence Reliability Diagram | https://skillware.site/bench/exhibits?id=fig27 | | fig28 | Figure 28 | Part III | Tier 1 Unit Tool Dispatch Telemetry | https://skillware.site/bench/exhibits?id=fig28 | | fig29 | Figure 29 | Part III | Tier 2 Context Injection Robustness | https://skillware.site/bench/exhibits?id=fig29 | | fig30 | Figure 30 | Part III | Tier 3 Adversarial Perturbation Response | https://skillware.site/bench/exhibits?id=fig30 | | fig31 | Figure 31 | Part VI | Retry Mass Distribution & Exponential Backoff | https://skillware.site/bench/exhibits?id=fig31 | | fig32 | Figure 32 | Part III | Tier 4 Complex Chained Workflow Critical Path | https://skillware.site/bench/exhibits?id=fig32 | | fig33 | Figure 33 | Part VI | 368-Layer Call Survival Flow | https://skillware.site/bench/exhibits?id=fig33 | | fig34 | Figure 34 | Part VI | Physical State Audit Timeline & Egress Telemetry | https://skillware.site/bench/exhibits?id=fig34 | | tab01 | Table 1 | Part II | 23-Model Fleet Master Leaderboard | https://skillware.site/bench/exhibits?id=tab01 | | tab02 | Table 2 | Part IV | Token Expenditure by Model & Crucible | https://skillware.site/bench/exhibits?id=tab02 | | tab03 | Table 3 | Part IV | Direct API Cost Waterfall per Episode | https://skillware.site/bench/exhibits?id=tab03 | | tab04 | Table 4 | Part IV | Economic Savings from Compact Model Substitution | https://skillware.site/bench/exhibits?id=tab04 | | tab05 | Table 5 | Part V | JIT Context Excision Memory Benchmark | https://skillware.site/bench/exhibits?id=tab05 | | tab06 | Table 6 | Part VI | Adversarial Telemetry Across Attack Vectors | https://skillware.site/bench/exhibits?id=tab06 | | tab07 | Table 7 | Part II | Four-Axis Master Operational Summary | https://skillware.site/bench/exhibits?id=tab07 | | tab09 | Table 9 | Part IV | Line-Item Token & Direct Cost Decomposition | https://skillware.site/bench/exhibits?id=tab09 | | tab10 | Table 10 | Part V | Statistical Reproducibility Across Crucibles | https://skillware.site/bench/exhibits?id=tab10 | ## Provider adapters (model-agnostic — add providers without rewriting skills) - SkillLoader.to_gemini_tool(bundle), to_claude_tool, to_openai_tool, to_deepseek_tool, to_bedrock_tool(bundle), to_ollama_prompt(bundle) - OpenAI-compatible hosts (Groq, OpenRouter, Mistral, Together, vLLM, LiteLLM proxy, etc.): use to_openai_tool() with host base_url — see https://skillware.site/documentation#openai-compatible - Gemini: match function_call.name via SkillLoader._sanitize_gemini_tool_name(manifest["name"]) - Pass bundle["instructions"] as system context; on tool call → skill.execute(args) - Documented today: Gemini, Claude, OpenAI, DeepSeek, Ollama, Bedrock (v0.5.6+), OpenAI-compatible — same skill manifest for any supported host Full usage guides: https://skillware.site/documentation Install extras: https://skillware.site/documentation#install-extras