Skip to content

ecological-codes/user-prefs

Repository files navigation

user-prefs

Platform-wide preferences and companion files for harness+model in agentic workflows.

Companion repository to ecological-codes/prompteng.

Introduction

Platform-wide self-instructions, System Prompt (SP) compensation guards, behavioral discipline rules, and session tooling for Claude.ai and Claude Code.

Designed to compensate for known deficiencies in Anthropic's SP for Sonnet 4.6 and Opus 4.7 — see sp-compensations-map.md for the full mapping.

Core mechanism: agent.md loads prompteng-SKILL.md every session. For Claude environments and models, claude-sp-guards.md and agent-prompt-discipline.md must also load every session because without them, agents tend to exhibit maladaptive behaviours and SP deficiencies go unchecked.

Set of Files

File Purpose
agent.md Reduces token usage by 60% to 70% by introducing file registry, re-read warning, integrity checks, memory precedence. Loads prompteng-SKILL.md for sanetizing given prompts. Supersedes archived/claude.md.
claude-sp-guards.md System Prompt (SP) compensation companion for Claude environments, provides memory conflict surfacing, canonization, hygiene, credential protocol. Load every session, required as long as Sonnet and Opus SP deficiencies persist. Adapt to other models as required.
agent-prompt-discipline.md Behavioral discipline rules for agents and subagents — assumption-surfacing, surgical edits, minimum viable output. Load every session — without it agents may exhibit maladaptive behaviours. Inspired by Karpathy Style Skills.
opus-thinking-mode.md Opus adaptive thinking configuration. Perishable — model-version-specific. Load for Opus sessions only.
sp-compensations-map.md Analysis — maps agent.md §1–§3 for compensating specific SP deficiencies in Sonnet 4.6 and Opus 4.7.
git-init-session.sh Session credential management script. Exports Personal Access Token (PAT) in .pat format to env var via file-upload + bash-pipe pattern; never writes to disk. Do not save PAT as .md or .txt file as they can get read directly into Context Window. This can be problematic and a security risk if "Session Memory" is enabled in the AI Platform.
memory-enablement-checklist.md Memory hygiene checklist across the four-tier model (Short-Term, Long-Term, Selective, Latent). For parallel agentic workflows.
trusted-hosts.md Project-wide fine-grained egress allow-list for bash_tool and outbound URL calls. Tries to mitigate supply-chain injection via compromised URLs.
archived/claude.md claude.md v1.6.4 — superseded by agent.md v2.0.0. Retained for reference.

Install

Every session (mandatory):

  • agent.md: paste its contents into Settings > General > Personal Preferences in Claude.ai or Clode Code, or adapt it to your specific harness.
  • agent-prompt-discipline.md: Better to paste this into Personal Preferences with agent.md. Required to prevent maladaptive agent behaviour.

As and When Needed:

  • claude-sp-guards.md: Don't merely add it to Skill Directory, it will under-trigger. Instead, add to Project Knowledge in each project or paste into Personal Preferences, when using any Claude enviornment or models. Otherwise required until Anthropic resolves Sonnet / Opus SP deficiencies.
  • opus-thinking-mode.md: upload to Project Knowledge or paste into session for Opus sessions only.
  • trusted-hosts.md: Add to Project Knowledge or upload into Context Window when needed.
  • git-init-session.sh: upload to Project Knowledge or upload into Context Window; run via source git-init-session.sh <PAT> for the PAT scoped git repo.

Quickstart

  • After installing, inititializes every session with first message.
  • Verify: Registry table appears with BLAKE3 + token cost for loaded files. If missing → re-upload + retry.

Peer Skills

Install into your Skill Directory, and load on demand when task requires:

  • prompteng — session init, security rules, 7-part prompt framework, persistence
  • captureng — session-knowledge capture, CHECKPOINT mode
  • packageng.skill file validation + packaging
  • safe-skill-creator — skill design + iteration

Reference

License

See LICENSE. (C) Copyright 2026 - Sameer Khan - Various and Several Rights Reserved.


README.md v1.6.1 - Human Approved

About

Examples of "platform-wide" preferences for harness+model in agentic workflows.

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors

Languages