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.

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

Mode Use when Core discipline
debug failing test, bug, regression reproduce first, one hypothesis at a time
fix small known urgent change smallest diff, no opportunistic cleanup
review security/auth/payments, reviewing a diff no approval without file/line evidence
test-first behavior change, tests practical see RED before implementing, never fake green
refactor cleanup, simplify, rename preserve behavior, trace call sites before deleting
optimize performance work, hot path measure first, baseline before/after
migrate schema/data/infra change rollback path before touching state
upgrade dependency or version bump read breaking changes, no blind search-replace
integrate calling an external API/service read the contract, handle the error paths
spike prototype, PoC, unknown library isolate, end with a verdict
unstuck repeated failures, thrashing stop editing, summarize evidence, narrow hypotheses

Always — in every mode:

  • Verify by running the real check (test, build, repro); never by re-reading.
  • Never report a result you didn't run; never weaken, delete, or game a test to go green.
  • No destructive commands (force push, reset --hard, drop, rm -rf) without explicit scope.

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

Agent Install
Hermes Drop skills/coding-posture/ into ~/.hermes/skills/ (auto-discovered). URL/hub install: see the Hermes skills docs.
Pi pi install git:github.com/alexei-led/coding-posture
Cursor / any SKILL.md agent Copy skills/coding-posture/ into the agent's skills dir.

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:

Claim Evidence strength Sources
A checklist shifts behavior via in-context conditioning Grounded substrate, not a full theory CoT (Wei 2022), induction heads (Olsson 2022), iteration heads (NeurIPS 2024), ICL/CoT (ICML 2024)
Procedures beat personas Supported guideline, not a law personas don't help (Zheng 2024), self-consistency (2022)
The model self-selects the mode Holds for strong models, not universal Route-to-Reason (2025)
The checklists target the right levers Highest-evidence levers + documented failures self-debug (Chen 2023), self-correction limits (Huang 2023), context-first (2026), verifier gaming (2026)
The skill helps +15pp in the eval — run it yourself eval/

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:

Layer What goes here Why
Always-on — CLAUDE.md / AGENTS.md the invariants (always-on-snippet.md) must fire every turn; a conditional skill can miss-activate
Conditional — this skill the per-task mode checklists loaded only when relevant; too many to keep always-on

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

Try it

Open source — clone it, drop the skill into your agent, and watch how it changes the work. In the behavioral eval (eval/) the skill scores 85% vs 70% without it (+15pp), with the urgent-auth case going 4/4 vs 2/4. Run the eval on your own model and tasks, open an issue with what you find, or add a mode — contributions welcome.