Coding Posture

Task-aware working modes for coding agents. One SKILL.md file: before non-trivial work, the agent picks a mode — debug, fix, review, test-first, refactor, optimize, migrate, upgrade, integrate, spike, unstuck — and follows its checklist.

It exists to stop agents from behaving like optimistic elevators with write access: thrashing on a stuck bug, faking green tests, skipping reproduction, running destructive commands, migrating without a rollback.

Coding-posture owns the general execution procedures. It works on its own; specialist skills are optional detail, not a second required checklist. See 0.6.0 release notes for the behavior changes.

Works with Hermes · Claude Code · Codex · Cursor · Pi — any SKILL.md-compatible agent.

How the agent uses it

flowchart TD
    Task["Non-trivial coding task"] --> Pick{"Pick a mode that fits<br/>(debug, fix, review, migrate, …)"}
    Pick -->|no match| Normal["Proceed normally"]
    Pick -->|match| Check["Follow the mode's checklist"]
    Check --> Loop["gather context → localize → smallest change<br/>→ run the real check → read output"]
    Loop --> V{"Verified by a real run?"}
    V -->|no| Loop
    V -->|yes| Done(["Done"])
    Guard["Always-on invariants:<br/>no fake green · no destructive commands · don't game tests"] -.->|guards every step| Check
    Guard -.-> Loop

Modes

Always — in every mode:

• Verify by running the real check (test, build, repro); never by re-reading.
• Report checks as passed, failed, or not run with a reason. A failing gate or blocker is not successful verification.
• Never hide failures or weaken coverage to improve a result; redundant tests may be removed when coverage is preserved.
• Stay within authorized scope, preserve unrelated changes, and respect read-only requests. Destructive actions and external sharing of private code require scoped authorization.

Install

Install as a plugin (one command, updates in place) or drop the skill in directly. It's a standard SKILL.md skill, so it works unmodified across compatible agents.

Claude Code — plugin

/plugin marketplace add alexei-led/coding-posture
/plugin install coding-posture@coding-posture

Update later with /plugin marketplace update coding-posture then /reload-plugins, or enable auto-update from the /plugin menu to refresh on startup.

Codex CLI — plugin

git clone [email protected]:alexei-led/coding-posture.git
codex   # then: /plugins → add the local marketplace in ./coding-posture → install coding-posture

Or point Codex straight at the skill in ~/.codex/config.json:

{ "skills": ["/abs/path/to/coding-posture/skills/coding-posture"] }

Update later with git pull.

Other agents

The agent activates the skill from its description when a coding task starts.

How it works — theory and evidence

A skill is just text placed in the model's context. A worked procedure acts as an _in-context demonstration_: the model conditions its next tokens on the shown trajectory, not just the final answer — the same basis as chain-of-thought. No complete mechanistic theory of frontier-model reasoning exists yet, so treat this as a grounded substrate, not a proof.

Each design claim, and how strong the evidence actually is:

Two deliberate consequences:

• Procedures, not personas. A persona ("act as an expert debugger") sets style; a procedure supplies structure the model can follow. So each mode is a checklist, not a character.
• Small on purpose. Instruction-following degrades as a prompt grows long and complex, so a short, followable procedure beats a long aspirational one. (The popular "~150–200 instructions" ceiling has no peer-reviewed source; the real effect is just degradation with length.)

Where this fits

Frontier agents already do the basics (reproduce bugs, run tests, small diffs), and a good CLAUDE.md/AGENTS.md repeats them. This skill earns its place only as the delta — the anti-instincts agents get wrong by default (stop thrashing, don't game the grader, roll back before migrating, measure before optimizing, read the API contract) — plus a mode catalog too large to keep always-on.

Split the rules by how often they must fire:

Anthropic's skill guidance is evaluation-driven: keep only what measurably closes a gap.

With cc-thingz

Use coding-posture for ordinary debugging, fixes, review execution, test-first work, refactoring, and optimization, including work on tests. Security-sensitive fixes stay implementation tasks with stronger evidence requirements; review does not silently make them findings-only.

Load cc-thingz's reviewing-code when its severity rubric, language checks, review depth, or report format is useful. Load improving-tests for test seams, mocks, assertions, characterization techniques, or framework and suite-speed detail. Language and toolchain guidance stays in cc-thingz. Those specialists do not need to be loaded for every coding task.

This release changes only coding-posture. Older cc-thingz installations can still advertise overlapping fix/TDD workflows; their routing must be narrowed separately. No companion skill is installed, renamed, or removed here. Existing mode names and plugin installation paths are unchanged.

Try it

Run the behavioral eval and execution fixtures on your own model and tasks. The earlier five-case result of 85% versus 70% is historical; it does not establish the effect of 0.6.0, its expanded cases, or newer models. Injected answer evals also do not prove skill activation or tool execution. Open an issue with reproducible results.