Open skills

We have opened the skills we use in our own work. Most of them are methods for reasoning honestly: auditing a tension, counting how many witnesses for a claim are really independent, closing loops by turning a named open item into the next build target, and a pre-registered gate for research campaigns. The rest are practical tools.

They are in one repository on GitHub, erdem-afk/skills, under the Apache License 2.0 and open to pull requests. Each skill is a folder with a SKILL.md, which Claude Code and other agents that load such folders read; the repository’s README says how to install one.

Method family

The harman gate and its engines. The Harman page explains what the engines are for and how far each has been checked; here they are by name.

• harman: the single gate for multi-lane research campaigns: it tests whether a campaign is warranted, freezes a held-out set, orders source-blind lanes and a one-shot transfer gate, and records closure and promotion receipts.
• pure-dialectic: judges the structure of one tension (a real asymmetric resolution, a symmetric pair, a category error, a bad infinity or a false dilemma) and leaves a receipt with an operational rule and a named residual.
• agentic-pure-dialectic-tree: audits a branching tree of tensions and gives every node its own verdict, never one verdict for the whole tree.
• freqdialectic-audit: counts how many of N witnesses for a claim are independent of a shared source, and restates the claim honestly.
• loop-closing: judges whether a residual is named, which channel it belongs to and whether a claimed closure was actually earned.
• dialectic-lifecycle: judges a long inquiry as a graph of stages that can stay open or be reopened by new evidence.
• reflexive-consolidation: finds which decision-changing signal a summary dropped, by rereading the source with independent lanes.
• begriff-scout: before a campaign spends its budget, tests whether its concepts are sharp enough to tell experiments apart.
• marx-filter: checks whether a Hegelian reading stays in thought or names the field of practice where it would be tested.
• storm-conductor: judges the delivery of one multi-lane run without writing product code.
• grand-conductor: judges a plan of several lanes or runs for order, custody, safety, trace integrity and independent witnesses.

Reasoning and writing

• recursive-descent-audit: audits a research line or a decision history level by level, and descends only where a node earns a child, until it grounds.
• real-loop-closing: for residuals that only a real-world act can close, replaces more desk analysis with the smallest pilot that produces the missing witness.
• logik: a discipline for the edge of an irreversible or one-shot act: it reads in advance how the plan could collapse, then rules fire, wait and close, or change the game.
• pergel: takes a treatment just applied to one example, completes it there, then opens it along checked orthogonal axes until the map saturates.
• due-diligence: checks whether a company, employer, counterparty or person is what it claims to be, by separating self-published sources from the few witnesses a fraud cannot fake.
• business-playbook: reads a firm as a polity: 19 governing questions for plans, org design, succession, risk and unit economics, distilled from Antony Jay’s Management and Machiavelli and the modern canon.
• benchmarking: honest, reproducible quality numbers for retrieval and speech models: the right metric, an independent fixture, a baseline and a stated confidence.
• pedagogical-writing: explanatory writing that teaches; it chooses the direct or the narrative register first.
• visual-explanation: explains with a drawing: single-file, print-friendly HTML and SVG pages in which every mark answers a reader’s question.
• explain-as: picks the format of an explanation: controlled-language English (ASD-STE100), a diagram, an HTML page or a narrated video.
• claude-5: rewrites a prompt, CLAUDE.md, skill or brief written for older Claude models in Claude 5 idiom, with a receipt of every line cut, moved or added.

Tools

• draw-io-diagrams: real .drawio files (flowcharts, UML, ERDs, C4, state machines) from a short text spec, with automatic layout, a linter and headless export.
• apple-glass: Apple-style glass surfaces on the web or natively, built to Apple’s rules and measured in a real browser.
• wiring-audit: finds unwired, dead or placeholder code with three parallel agents.
• checkpoint-capture: durable, dated checkpoint notes with Obsidian-style frontmatter, backlinks and lineage.
• medusa: runs one prompt as a checkpoint-backed headless loop across Codex CLI, Claude Code and Gemini CLI.
• performance: a process and memory inventory for a slow Mac that stops only what you pick and shows the counters before and after.
• tinker: an operating manual for the Tinker service and its Inkling models: sign-in, sampling text, images and audio, prices and billing.
• book-graph-extraction: turns books (PDF, scans, EPUB) into searchable Markdown and, if you want, a knowledge graph whose every edge carries a verified quote.

Before you rely on one

Most of these methods are at an early status, seed or candidate: they have been tested on a small sample, sometimes in a single run. Each method says in its own SKILL.md what was tested and what was not (its honest floor). Read it before you use the method, and do not treat any of them as proven.

Contributing

Pull requests and issues go to the repository; its CONTRIBUTING.md has the ground rules. A skill can tell an agent to run shell commands and the scripts in its folder, so read a skill before you install it.