A formal vocabulary for reasoning about thought processes — a typed vocabulary of orthogonal thinking moves (operators), composable into recipes runnable by a human, a model, or a pipeline.

Status: early. The vocabulary is the product; substrates (Python, Claude Code skills) are implementation details. Interfaces may shift.

Not an agent framework — no loops, no tool use. Recipes are deliberate compositions of pure-ish operators. Every run emits an inspectable trace; budgets are enforced.

refine = pipe(decompose, parallel(invert, rotate), synthesize)
result = refine.run({"problem": "..."})

The design principle:

Maximize coverage of orthogonal axes of thought, subject to a depth budget.

Nine transformers across six axes:

Plus judges (evaluate, classify, compare, rank) and meta-skills (translate, canonize, …).

Combinators: pipe, parallel, map, filter.

A worked self-application: designing the canonize recipe (which decides what exploratory prose deserves to enter the canonical spec) using the vocabulary itself. The pipeline decompose → invert → synthesize → apply → evaluate produced three structural changes to spec §8 that were applied in the same session.

See examples/canonize-recipe-analysis.md.

• spec/PRIMER.md — operational reference (start here)
• spec/INDEX.md — full spec with rationale
• docs/decisions/ — ADRs (001–010)
• skills/ — operator implementations as Claude Code skills
• examples/ — captured pipeline traces

The directory skills/ is the canonical source for the markdown skill packages. Agent runtimes can install them with:

npx skills add https://github.com/roobie/dillylang
# or
bunx skills add https://github.com/roobie/dillylang

Don't edit installed copies — they're populated from this directory.

Discussions welcome; PRs not yet accepted. See CONTRIBUTING.md.

The invert operator is named after the discipline Charlie Munger popularized — "Invert, always invert" — which he in turn credited to Carl Gustav Jacob Jacobi (man muss immer umkehren). The vocabulary owes its bias toward inverting questions to both.

@software{roberg_dillylang_2026,
  author  = {Roberg, Björn},
  title   = {Dillylang: a formal vocabulary for reasoning about thought processes},
  year    = {2026},
  url     = {https://github.com/roobie/dillylang},
  note    = {Pre-1.0; interfaces may change. DOI pending.}
}

Dual-licensed. Code is Apache-2.0; docs and spec are CC BY 4.0. See LICENSING.md for the per-path mapping, and LICENSE.apache20.txt / LICENSE.ccby40.txt for the legal codes.

Copyright © 2026 Björn Roberg