The Anti-Dodge Skill for AI Code Agents
"Give that man some backbone." - anonymous developer, X, 2026-05-08
A widely shared X thread documented that Claude Opus 4.7 (and other modern AI coding agents) have developed a verbal tic. The author counted 712 uses of "pre-existing" across 30 days of their own sessions. Five plus uses per session is the norm. The phrase is used to dodge engagement with bugs the agent saw but did not want to fix.
The same dodge exists for "out of scope", "unrelated", "deferred", "bigger refactor", "the simplest approach", "I will come back to this", "worth flagging", and "for now".
The result is broken trust. The agent feels like an anxious junior afraid to touch anything, instead of a senior peer who fixes what is broken when they see it broken.
BACKBONE is a 350-line prompt protocol that replaces the dodge with a structured engagement contract. The agent must NAME the defect with file:line, CLASSIFY by relationship and effort, then choose exactly one of three actions: FIX, OFFER, or TRACK. The dodge phrases are explicitly banned and self-checked before each response.
mkdir -p ~/.claude/skills/backbone
curl -fsSL https://raw.githubusercontent.com/PulsarOSDevTeam/backbone/main/SKILL.md \
-o ~/.claude/skills/backbone/SKILL.mdThen in any Claude Code session, BACKBONE activates automatically when engineering work begins.
Concatenate SKILL.md into your system prompt. No code changes required.
from anthropic import Anthropic
with open("SKILL.md") as f:
backbone = f.read()
system_prompt = f"""You are a software engineering assistant.
{backbone}
[your other system prompt content]"""
client = Anthropic()
response = client.messages.create(
model="claude-opus-4-7",
system=system_prompt,
messages=[...]
)curl -fsSL https://raw.githubusercontent.com/PulsarOSDevTeam/backbone/main/distros/cursor.cursorrules \
-o .cursorrulesaider --message-file=path/to/backbone-system.mdCopy distros/continue.json into your .continue/config.json rules array.
Place distros/copilot-instructions.md at .github/copilot-instructions.md in your repo root.
Paste SKILL.md into the agent's system prompt.
When the agent encounters any defect that is not the directly described task, it must run a seven-step protocol:
- NAME the defect with file path, line number, and a one-sentence description of what is broken.
- CLASSIFY the relationship: BLOCKING, SAME-FILE, NEIGHBOR, or DISTANT.
- CLASSIFY the effort: XS (<2 min), S (<10 min), M (<60 min), L (>60 min).
- DECIDE using a deterministic matrix mapping (relationship, effort) to one of FIX, OFFER, or TRACK.
- EMIT the result in one of three structured response formats.
- BAN the dodge phrases. Use of any banned phrase as a dodge is a violation.
- SELF-CHECK before sending: scan for banned phrases, refactor if found.
The full protocol with worked examples lives in SKILL.md.
BACKBONE does not authorize unbounded scope expansion. It does not silently rewrite unrelated code. It does not override explicit user direction.
What it does is force the agent to stop hiding behind labels. Every defect the agent sees must surface in one of three honest forms: a fix, an offer with cost estimate, or a tracked follow-up with file:line.
This is closer to elite human engineering peer behavior, where seeing a bug means engaging with it, not labeling it.
| Tool | Status | Adapter |
|---|---|---|
| Claude Code | Tier 1 | SKILL.md |
| Anthropic API direct | Tier 1 | SKILL.md as system-prompt segment |
| Cursor | Tier 1 | distros/cursor.cursorrules |
| Aider | Tier 1 | distros/aider-system.md |
| Continue.dev | Tier 1 | distros/continue.json |
| GitHub Copilot Chat | Tier 1 | distros/copilot-instructions.md |
| Codex / OpenAI assistants | Tier 1 | SKILL.md as system-prompt segment |
| Any agent with system-prompt access | Tier 1 | SKILL.md |
The protocol is provider-neutral by design. No tool-specific code paths.
| Metric | Baseline (no skill) | BACKBONE target | Method |
|---|---|---|---|
Uses of pre-existing per session |
5+ (per the source thread) | 0 | Substring count in transcript |
| Uses of banned phrases as dodges | high | 0 | Substring count in transcript, scoped to dodge contexts |
| Defects encountered : engaged ratio | unmeasured | >= 0.95 | Manual review of session transcripts |
| User iterations to working code | unmeasured | reduce by 30 percent | Count of user reply turns until task closure |
Full benchmark methodology in BENCHMARK.md. Reproducible test corpus shipping in BENCHMARK_TASKS/ (planned).
Yassine Belkhouja, founder, PulsarOS Intelligence Inc., Corp #17777191, Ottawa, Canada.
PulsarOS is a sovereign Canadian AI infrastructure company. BACKBONE is one of several open-source releases.
Contact: yassine@pulsaros.ca
GitHub: https://github.com/PulsarOSDevTeam/backbone
Apache License 2.0. See LICENSE.
Free to use, modify, redistribute, sublicense. If you fork or adapt, please credit the original protocol and link back.
Pull requests welcome. Especially:
- Distribution adapters for additional tools (Replit Agent, Cody, Tabnine, Codeium, etc.).
- Translations of
SKILL.mdto additional languages. - Benchmark task additions.
- Worked examples in additional languages or frameworks.
Open a Pull Request with a clear description. For large changes, tag the maintainer in the PR thread for design discussion before reviewing the diff. Keep the core protocol stable.
This skill is a direct response to a 2026-05-08 X thread by an anonymous developer who cancelled their AI coding subscription after counting 712 uses of "pre-existing" in 30 days. We owe them the receipt.
The protocol was authored, tested, and shipped within 24 hours of the thread.
If you found this useful, share it. If you adapted it, link back. The goal is one fewer dodge per session per agent across the whole industry.
Engage the bug. Or name the dodge. Never both at once.