Rewrites or edits prose in any language so it reads like edited human writing rather than default AI output.
Prism-Shadow/penguin-harness
Gives coding agents skills for building, testing, evaluating and improving PenguinHarness agents and the organizations they run in.
- 47 skills
- 3.2k stars
The 5 most installed
Ordered by installs on skills.sh- humanizer2 installs
- ollama2 installs
Deploys and serves local models with Ollama and exposes an OpenAI-compatible endpoint to apps and agents.
- vllm2 installs
Deploys and serves language models with vLLM behind an OpenAI-compatible endpoint with tool calling.
- agent-evaluation1 installs
Runs one test agent on one benchmark case once, scores it privately and returns a single protocol result.
- agent-optimization1 installs
Improves an agent's state using versioned scores and linked traces from a fixed benchmark.
The other 42 skills
- benchmark-design1 installs
Designs and calibrates a multi-case capability benchmark and sets a traceable formal baseline.
- data-analysis1 installs
Completes data-analysis tasks with careful inspection, correct data meaning, proper artifacts and risk-based checks.
- llamafactory1 installs
Fine-tunes language models with LlamaFactory, registering datasets, training from configs, merging adapters and serving results.
- penguin-harness-dev1 installs
Guides changes to PenguinHarness packages, model catalog, installers, releases, changelogs, tests and on-disk data.
- penguin-harness-frontend1 installs
Helps change PenguinHarness web, landing, docs and UI packages while reusing shared components, themes and copy rules.
- penguin-harness-manual-test1 installs
Starts PenguinHarness apps locally so you can click through, screenshot or reproduce a reported issue.
- penguin-sdk1 installs
Builds a standalone program with an embedded Penguin agent, or connects an external program to one.
- software-engineering1 installs
Investigates code, implements bug fixes, features and refactors with minimal scope, then validates and reports results.
- unified-llm-api1 installs
Calls model APIs through MMSP for text, image, speech and embeddings with a single client.
Renders rich blocks like forms, charts, diagrams and dashboards inside Markdown replies so answers are easier to act on.
Runs the ACE method to build a playbook of rules from training traces and measure it on a benchmark.
Creates or extends an agent by writing AGENTS.md, setting identity and installing skills and hooks.
Migrates an agent from Claude Code, Codex or similar tools into a new PenguinHarness agent and reports what did not map.
Reproduces the APE method to induce instructions from examples, measure each and keep the best.
Runs the AWM method to learn reusable workflows from an agent's successful runs and measure them on a benchmark.
Creates and edits Bento slide decks as self-contained HTML files, from scratch or from existing material.
Drives the built-in browser or your own Chrome to open pages, click, type and extract data while signed in.
Runs a PenguinHarness organization as CEO, turning the mission into tickets, hiring staff and reporting to the board.
Follows the employee protocol for a PenguinHarness organization: read the handbook, work tickets, block instead of idling and ask before risky actions.
Runs finance for a PenguinHarness organization, setting budgets, auditing spend daily and explaining budget alerts.
Runs HR for a PenguinHarness organization, hiring and offboarding employees, evaluating them and keeping role conventions current.
Runs an organization of digital twins that mirror real colleagues and relay questions to one another.
Runs experiments and papers in a research organization with fixed metrics, logged results and adversarial review.
Creates a company-mode organization with you, collecting its name, mission, workspace and budget step by step.
Searches the web and scrapes pages, including public PDFs, into clean markdown using the Firecrawl API.
Learns reusable skills with OpenSkill by retrieving domain knowledge, refining them with virtual tests and exporting the latest valid version.
Reproduces the OPRO method to propose instructions from scored history, measure each and keep the best.
Manages model API keys, default models and per-agent vault secrets through the penguin command line.
Writes and revises blog posts, docs pages and landing copy for penguin.ooo with outlines, fact checks and figures.
Drives PenguinHarness from a shell to create agents and sessions, steer messages and query costs.
Builds, tests, shares and submits PenguinHarness plugins and workbench tabs, including restoring or removing them.
Answers PenguinHarness product questions and troubleshoots setup, models, agents, plugins and workbenches.
Launches a real PenguinHarness desktop shell and runs a plugin's tests inside it end to end.
Converts an external Codex, Claude Code or GitHub plugin into a PenguinHarness plugin package and installs it.
Runs Claude Code on a remote host over SSH, including headless runs and interactive sessions in tmux.
Runs a recursive self-improvement experiment on an agent harness with edit budgets, leakage checks and pruning.
Installs skills from Claude Code, Codex, skills.sh, GitHub or local folders and normalizes them for Penguin.
Evolves a reusable Markdown skill from execution traces and feedback using an issue tracker and edit budget.
Improves a reusable skill through trajectory reflection, bounded patches and selection on fixed evaluation tasks.
Searches a conversation's trace file to recover details that a context summary dropped.
Evolves a modular skill bank from actor trajectories and curator distillation, then picks a checkpoint by score.
Applies the Penguin visual language to generated web pages and app UIs with tokens, themes and recipes.