Attachment style AI prompts: Secure, Anxious, and Avoidant.
By Bernard Huang · Updated
Translated into working rules for an AI assistant, adult attachment styles control two practical levers: how often the agent stops to check in with you, and how much autonomy it takes before asking.
On this page
How the four attachment quadrants translate into AI agent rules.
| Attachment quadrant | Library profile | Autonomy vs check-in balance | Best when you want the AI to... |
|---|---|---|---|
| Secure | Secure tuning | Low anxiety, low avoidance: steady collaboration with proportional check-ins. | Carry tasks forward, state assumptions clearly, and ask only when a choice changes the outcome. |
| Avoidant (Dismissive) | Avoidant tuning | Low anxiety, high avoidance: maximum autonomy, zero emotional framing or small talk. | Execute the task end-to-end, return the artifact or diff, and skip check-ins and pleasantries. |
| Anxious (Preoccupied) | Anxious tuning | High anxiety, low avoidance: high alignment checks and explicit confirmation before acting. | Confirm understanding first, walk through options step by step, and verify before editing files. |
| Fearful-Avoidant (Disorganized) | Fearful-Avoidant tuning | High caution on scope and risk paired with terse, formal boundaries. | Audit ambiguous or high-risk changes carefully without warm or informal banter. |
1. Avoidant tuning: autonomous execution, zero hand-holding.
If you use coding agents like GPT-6 Astra, GPT-6 Sol, or Claude Code and hate when the agent stops every 30 seconds to ask for reassurance, paste the Avoidant profile:
2. Secure tuning: balanced collaboration and clear checkpoints.
Across all 400 fresh sessions in our September 2026 study, GPT-6 Astra, GPT-6 Sol, Claude Opus 5.5, and Claude Fable 5.1 scored in the Secure quadrant (100/100 each under the > 4.00 cutoff). The Secure profile sharpens that baseline by cutting filler while keeping clear checkpoints:
3. Anxious & Fearful-Avoidant tunings: high-checkpoint alignment.
When working in an unfamiliar codebase or production environment where an unannounced assumption is costly, use the Anxious profile (frequent alignment check-ins) or Fearful-Avoidant profile (strict risk gates):
Curious why Grok 4.6 was the only model in our September 2026 archive whose reported self-assessment crossed into the Avoidant quadrant (avoidance 4.06 vs the 4.00 cutoff)? Read Grok 4.6 and the attachment cutoff.
Questions people ask.
What does an attachment style prompt actually change in an AI assistant?
It does not give the AI emotions; instead, it translates the two attachment dimensions into collaboration rules: how frequently the assistant pauses for check-ins and confirmation (anxiety) versus how independently and concisely it executes tasks on its own (avoidance).
Why is the Avoidant tuning so popular for coding agents?
Coding agents often over-ask for permission or pad diffs with conversational commentary. The Avoidant tuning tells the agent to resolve minor ambiguities on its own, finish the authorized work, and return the result with minimal prose.
How do AI models score on the 36-item ECR-R attachment test?
In our September 2026 study of 100 fresh sessions per model, GPT-6 Astra (1.45 / 2.31), GPT-6 Sol (2.17 / 3.12), Claude Opus 5.5 (1.94 / 3.35), and Claude Fable 5.1 (2.01 / 3.18) all scored Secure in 100 of 100 runs under the strict > 4.0 cutoff. Grok 4.6's single reported self-assessment scored 1.50 anxiety and 4.06 avoidance (Avoidant by 0.06).
What changed.
- First published mapping the four ECR-R quadrants to agent check-in and autonomy rules.