argot (n.) — the specialized working language of a craft guild.
A minimal primitive command language for senior technical work. Seven verbs, stable for a decade; skills, prompts, models, and harnesses are interchangeable implementations underneath.
understand · design · plan · build · verify · review · share
Today's agent interfaces expose implementation detail: hundreds of slash commands, skill piles, personas, model toggles. Every durable command language humans have built — Unix syscalls, git porcelain, CAD verbs, Photoshop tools, Polya/OODA/PDCA — is instead 4–10 generative verbs over a frozen object model, with composition kept outside the verbs. Agent systems are converging on the same set by independent evolution. argot names that convergence and freezes it as a contract.
The durability law that selects what belongs here:
Abstractions die in proportion to how much model behavior they encode, and survive in proportion to how much task intent they encode.
So argot contains only task intent. Everything model-shaped (routing, effort, panels, budgets) lives in resolver policy below the interface.
| Layer | Contents | Stability |
|---|---|---|
| Verbs (porcelain) | 7 signatures + operator algebra + channels | frozen — SPEC.md |
| Artifacts (object model) | Brief, Design, Plan, Change, Verdict, Findings, Communication | frozen — artifacts/ |
| Resolver (policy) | (verb, target, standards, depth) → implementation; model routing | evolves — resolver/ |
| Implementations (plumbing) | skills, playbooks, subagents, tools, models | churns freely |
| Verb skills | thin routers, standard skills layout | skills/ |
One deliberate omission defines the design: decide is not a verb. The agent frames decisions
(design emits options + tradeoffs + recommendation) and records them (share); the human decides
at a gate (⊸) in the composition algebra. Verdicts, never actions, cross that line.
understand src/auth --deep # → Brief (cited, confidence-tagged)
design "rate limiting for the public API" # → Design (options + tradeoffs + recommendation)
plan design:0142 # → Plan (tasks + acceptance criteria) ⊸ your approval
build plan:0143 # → Change (diff + decision trail)
verify change:0144 # → Verdict {verified|refuted|inconclusive} + evidence
review change:0144 --against security # → Findings (fresh context; never auto-applied)
share change:0144 --as pr # → Communication (outward gate)
Workflows are compositions, not new verbs:
feature = understand → design ⊸ plan ⊸ (build → verify)* → review → share
bugfix = understand(failure) → verify(repro) → build → verify → share
audit = understand(surface) → review --panel --against threat-model → verify(findings) → share
incident = (understand → verify → build?) --loop until-stable → share(postmortem)
Familiar words like diagnose, refine, watch are aliases that expand visibly to verb
expressions — see aliases.md. The expansion prints on use; the algebra teaches itself.
Standard skills layout (agentskills.io) — pick your route:
# skills CLI (Claude Code, Codex, Cursor, and friends)
npx skills add jschilli/argot -g
# or as a Claude Code plugin
/plugin marketplace add jschilli/argot
/plugin install argot
# or dev-mode from a checkout (symlinks track your edits)
ln -s "$(pwd)/skills/"* ~/.claude/skills/Then /understand, /design, /plan, /build, /verify, /review, /share are live.
Notes: the verb skills are thin routers — they resolve into whatever implementation skills you
already have (see resolver/ROUTING.md) and enforce the artifact contract.
Artifacts land in .argot/ of the working repo (markdown + JSON sidecar); gitignore it where you
dogfood. If a verb name collides with an existing skill of yours, keep yours — argot's verb can
route into it.
A new verb is admitted only if it cannot be expressed as verb + target + standards + modifier
(the admission test, SPEC.md §6). This test is what keeps the language at seven for a
decade. Implementations need no permission to change; the spec changes by versioned PR only.
Because that's what this is: the private working language of a craft — compact, precise, learned once, spoken for a career. The verbs are the argot; the repo just writes it down.
The full research report — 4 exemplar skill-repo harvests, 9 harness surveys, 9 framework surveys, human-analogue studies, agent-trace literature, and a 230-skill corpus clustering — lives at docs/design-rationale.md.