ajitta/dialectic-skills

English | 한국어 | 日本語 | 简体中文
Website: ajitta.github.io/dialectic-skills
Ten reasoning-and-review skills in one plugin, written to run unchanged on any assistant that loads Agent Skills: Claude (web, desktop, mobile, Claude Code), Codex and ChatGPT. The package is instructions only (Markdown, YAML, JSON), so every skill also works in a chat with no tools. Skill files are in English; each skill replies in the language of your message. On a phone, type just dia: a menu skill (dia-menu, new in 1.4.0) lists the skills and starts one by number, by the first letters of its name, or by what you ask for.
| Skill | What it does | How it starts |
| dia-confirm | Before work starts: restates the goal, discloses material assumptions, lists only the decisions you must make, then waits for your approval. | by name only |
| dia-elenchus | Plato's elenchus: "what is X?", premises agreed one question at a time, contradictions shown, each revision is your new idea. Usually ends in aporia; gives no advice. | by name, or by asking for a Socratic dialogue |
| dia-brainstorm | Modern Socratic questioning one question at a time, then your options plus up to three labeled ones, convergence by criteria, a verdict and a brief. Never implements. | by name, or by asking for a Socratic brainstorm |
| dia-reductio | Strict reductio ad absurdum in one pass: labeled premises, a Q and not-Q derivation with a rule per line, graded strength, options and a record. | by name, or in words |
| dia-negative | Internal critique (narrower than Adorno's immanent critique) in one pass: the target's own standard, strict gaps with cases and "what escapes", the operative purpose with a falsifier, options without forced synthesis. | by name, or in words |
| dia-warrant | Toulmin's argument layout with Walton's critical questions in one pass: data, warrant (the unstated one supplied and labeled), backing, qualifier and rebuttal, each question tagged with the part it tests, then the weakest link. Never rules on whether the claim is true. | by name, or in words |
| dia-conjecture | Lakatos's proofs and refutations as a dialogue on an "all / always / never" claim you hold: your argument split into steps, one counterexample per message with its check, you rule on each case and pick the repair. | by name, or in words |
| dia-counterplan | Mason's dialectical inquiry on a plan or decision: the key assumptions, a plausible contrary for each, a counterplan on the same facts, then you say which assumptions you now hold. Picks no winner. | by name, or in words |
| dia-refine | One review pass over the last answer or a draft: factual errors, logical leaps, missing requirements, repetition. Fixes only confirmed defects. | first word dia-refine, or by name |
| dia-iterate | Repeated review rounds, each gated by external verification; stops on regression, zero defects, verification exhausted, fixed point or --max N (default 3). | first word dia-iterate, or by name only |
| dia-menu | Not a method: a short entry point. dia shows a numbered menu, dia 4 … or dia red … starts a skill by number or first letters, and dia picks the skill that fits and says which. Never starts dia-confirm or dia-iterate on its own. | first word dia |