Give AI coding agents a backbone.
Backbone removes reflexive agreement, praise, apology theater, and verbose task restatement. Agents acknowledge the task or their exact error briefly, test claims before agreeing, and spend tokens on evidence, implementation, and verification.
npx -y github:ThinkPool-io/backbone install --allInstall for one or more runtimes:
npx -y github:ThinkPool-io/backbone install --codex
npx -y github:ThinkPool-io/backbone install --claude
npx -y github:ThinkPool-io/backbone install --hermesWithout runtime flags, Backbone detects installed agents and asks which ones to modify. Use --dry-run to inspect the planned changes first.
Restart selected agents after installation.
Backbone is opt-in. The installer adds one marker-fenced instruction block and one owned skill directory for each selected runtime:
| Runtime | Always-on instructions | Skill |
|---|---|---|
| Codex | $CODEX_HOME/AGENTS.md or ~/.codex/AGENTS.md, plus an existing active AGENTS.override.md |
~/.agents/skills/backbone |
| Claude Code | $CLAUDE_CONFIG_DIR/CLAUDE.md or ~/.claude/CLAUDE.md |
Claude’s user skills directory |
| Hermes | $HERMES_HOME/SOUL.md or ~/.hermes/SOUL.md |
Hermes’ user skills directory |
Existing instructions are preserved. Reinstalling updates the managed block without creating duplicates. Backbone refuses to overwrite a skill directory it does not own.
npx -y github:ThinkPool-io/backbone status --all
npx -y github:ThinkPool-io/backbone uninstall --allUninstall removes only Backbone’s marker-fenced instructions and installer-owned skill directories. Unowned files are preserved.
- Start work directly.
- Name your own exact error once, then repair it.
- Do not praise users for catching mistakes.
- Do not agree until the claim survives an evidence and counterargument check.
- Do not become contrarian for sport.
- Cut canned rhetoric by function instead of playing phrase whack-a-mole.
- Prefer literal language while preserving exact technical vocabulary.
- Ask only when a safe default would materially change the outcome or authority.
- Continue until completion, a genuine blocker, or a user-set boundary.
- Match confidence to observed evidence.
- Preserve technical precision.
- Expand when safety, sequencing, or ambiguity requires it.
- Stop when the useful answer is complete.
The portable prompt lives in skills/backbone/references/standalone-prompt.md. The installable skill lives in skills/backbone/SKILL.md.
The installer has no runtime dependencies and makes no network requests. It edits only the selected global instruction files and Backbone skill directories. Use --dry-run before installing if you want to inspect every target.
npm test
npm run checkBackbone requires Node.js 18 or newer.
MIT