🇬🇧 English · 🇷🇺 Русский
A set of working postures — each one sets an agent's operating mode for a task. They answer questions like How much autonomy? Speed or quality? Do it myself or delegate? Minimal change or bold bets?
Most are chosen deliberately — invoke /surgeon or name the posture in your prompt, and switch when the situation changes. Three engage proactively and say so in their own descriptions: clarifier fires before acting on ambiguity, skeptic before a commit, archivist across long sessions. The rest can be pulled in by you, or surface automatically when a task clearly matches one. If more than one surfaces for a task, pick the single best match — only one posture should bind at a time (you can still layer a style posture like surgeon or archivist onto it). Each posture is intentionally opinionated: it optimizes for one thing and says out loud what it deliberately skips. That tradeoff is the point.
Install each <posture>/ directory into your agent's skills location, then invoke by name.
- Claude Code: copy the posture directories into
~/.claude/skills/(or a project.claude/skills/), then run/surgeon(or name the posture in your prompt). - Codex: place under
~/.agents/skills/. - Other agents: these are plain
SKILL.mdfiles with standardname/descriptionfrontmatter — drop them wherever your agent discovers skills.
Invoke one per task, or layer a style posture onto a doing one (see Combining and contradicting).
| Situation | Posture |
|---|---|
| You want analysis or a recommendation — no changes | advisor |
| Stress-test a plan or change before you commit | skeptic |
| Throwaway script or a hard time-box | sprinter |
| Library code, public API, or shared dependency | perfectionist |
| New module / prototype where structure is the bet | trailblazer |
| Hotfix or fragile legacy code — touch as little as possible | surgeon |
| Existing code needs restructuring without changing behavior | refactorer |
| Short, well-scoped task — just do it, no machinery | solo-operator |
| Long session with parallelizable work — delegate | orchestrator |
| Trusted goal you can walk away from, end-to-end | autonomous-executor |
| Task is ambiguous — resolve intent before acting | clarifier |
| You want to stay in the loop, step by step | pair-partner |
| Long project — must hand off context cleanly | archivist |
Autonomy — who drives each step:
pair-partner / clarifier (you drive, or resolve ambiguity first) → autonomous-executor (agent drives, you leave)
Speed ↔ Quality — what you optimize for:
sprinter (fast, disposable) ↔ perfectionist (exhaustive, polished)
Delegation — do vs. coordinate:
solo-operator (do it all directly) → orchestrator (delegate to subagents)
Disposition — constructive vs. adversarial:
advisor (reason and recommend) ↔ skeptic (attack and break)
Footprint — how much you change, on what ground:
| existing code | empty ground | |
|---|---|---|
| touch little | surgeon |
sprinter ¹ |
| restructure | refactorer |
trailblazer |
¹ sprinter's home axis is speed — it appears here because disposable code is structurally minimal, not because the task is small.
archivist is a style layer — documentation-first — that layers onto any of the above.
Most postures are chosen one-per-task, but the style postures layer cleanly: run an autonomous-executor task with surgeon discipline, or keep archivist notes through a trailblazer spike — name both in your invocation.
Several pairs are near-opposites (sprinter ↔ perfectionist, advisor ↔ skeptic, solo-operator ↔ orchestrator, surgeon ↔ refactorer, clarifier ↔ autonomous-executor). Invoking both is a contradiction; pick the one your situation calls for.