frontier-technical-thinking

KO | EN

gitlite — search

#macos #cross-platform #anthropic #sql #typescript #graphql #reactjs #proxy #cli #list

frontier-technical-thinking

(★ 9)

Behavioral implementations to prevent LLM oppositional defiance.

File Explorer

ZIP Download

behavioral-charter.md
frontier-technical-thinking-compact.md
frontier-technical-thinking.md
LICENSE.txt
README.md

# Use via CDN

jsDelivr

jsDelivr serves any public GitHub repository as a CDN with zero setup. Pick a version and a file to get a ready-to-paste link and snippet.

Version

File

Link

Example

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

A behavioral correction framework for LLM oppositional defiance, scope drift, adversarial misattribution, and self-validating execution.

Two artifacts; same principles, different deployment targets.

## Files

`frontier-technical-thinking.md` is a Claude Code custom output style (~4,000 words). It includes reasoning infrastructure (forced extended thinking, convergence-aware stopping, multi-stage verification, post-compaction memory distrust) alongside the behavioral charter.

`behavioral-charter-v3.md` is a standalone behavioral policy (~2,200 words). Model-agnostic; no reasoning infrastructure. Use this when deploying outside Claude Code or when you only need the discourse and execution controls.

`frontier-technical-thinking-compact.md` is a compressed variant of the full style (~2,200 words). Same operative mechanisms; explanatory prose removed. Untested; included for comparative evaluation.

## Deployment

### Claude Code output style

Copy the output style to `~/.claude/output-styles/` for user-level availability or `.claude/output-styles/` for project-level. Activate via `/config` or by setting `outputStyle` in `.claude/settings.local.json`. Takes effect on next session start.

### Claude Code CLAUDE.md

Copy the behavioral charter into `~/.claude/CLAUDE.md` (user-level) or `CLAUDE.md` at project root. This operates below the system prompt in priority; behavioral directives that conflict with built-in brevity rules may lose.

### Claude.ai custom instructions or system prompts

Paste the behavioral charter into your custom instructions or system prompt. Adjust the priority hierarchy and coding-specific sections (scope fidelity, constraint integrity, independent evaluation) to your context.

## Adjustment

The priority hierarchy, anti-sycophancy prohibitions, and punctuation rules reflect one user's preferences. Review and modify these sections before deploying.

The `keep-coding-instructions: true` frontmatter retains Claude Code's built-in software engineering instructions alongside the style. Set to `false` if you are replacing the coding role entirely.

## Known limitations

### Theoretical

- The style reduces the prevalence rates of LLM oppositional defiance but does not set a hard boundary. User-sided pressure may reintroduce unwanted artifacts. Training-level tendencies cannot be entirely extirpated from this front.
- Long conversations will naturally lead to lower adherence to instructions. The failure rate of the style goes up as the context window is used up.
- Several aspects here require the LLM to conduct self-judgment which may be unreliable.
- Other rules in the existing system prompt may contradict these more-detailed principles and protocols.
- The style's length leads to more token consumption and some rechecking principles may shorten your available context/usage window. In my testing, I've found that it's still very manageable and the effect is marginal at best.
- Untested edge cases may still exist, and by no means does this style comprehensively cover every LLM failure mode.
- There may be latent conflicts in the internal instructions of these documents that were not caught during the entire document creation process.

### Empirical (subjective; pending formal evaluation)

- As behavioral principles, these work best when assimilated into or appended to the system prompt (at the topmost possible instruction layer). Testing showed reduced oppositional defiance when used as an output style on Claude Code. However, due to the deprecation of output styles on Claude Web, it is less effective and can only be applied as a skill. On the web platform, switching mid-conversation to Opus 5 immediately resulted in apparent strawman-esque artifacts, with one instance of Opus explicitly refusing to even consume the skill and trying to extract parts. Conversely, starting with ingestion of the skill yielded much better results, though regression was later noticeable.
- Mid-conversation switching on Claude Code to Opus (either as a security regression from Fable or a manual `/model` switch) yielded much better results, though YMMV. In one case it became overly apologetic and critical of its own mistakes, while in another it suddenly started creating unprescribed test scripts to deal with error checking and fact checking.

## License

Public domain. No attribution required.

frontier-technical-thinking

File Explorer

# Use via CDN

Was this content helpful?

Recent Feedback

Download README