Skip to content

TAUT v0.13 — first public release

Choose a tag to compare

View all tags
@0point9bar 0point9bar released this 16 Apr 20:40
v0.13
0ff1e98

Headline

A communication-style system prompt for coding agents that compresses prose output by 80% on average across 8 different agent CLIs while keeping a professional senior-engineer register.

Agent Δ % vs baseline Compliance
gemini −96.0 % 100 %
droid −85.8 % 100 %
pi −79.0 % 100 %
openclaw −78.6 % 100 %
codex −77.4 % 93 %
claude −76.2 % 100 %
hermes −75.2 % 87 % *
cursor-agent −71.3 % 93 %
TOTAL −80.0 % avg 96.7 %

* hermes ceiling at 87% is harness-bounded — its CLI binary forcibly emits diff views and session_id: trailers no prompt can suppress. See EVOLUTION.md §8.

Beats caveman

Caveman's published average is ~65% reduction. TAUT lands at 80%. Beat by 15 percentage points while keeping a professional senior-engineer register (no persona collapse, no character roleplay, no "ooga booga" responses).

What's inside

  • TAUT.md — the prompt. Drop into your agent's global instruction file.
  • AGENT-LOCATIONS.md — per-agent deploy paths + one-shot bash script.
  • README.md — entry point + quick install.
  • PHILOSOPHY.md — design rationale, ML grounding (Constitutional AI, InstructGPT, Sycophancy paper, Role-Play LLMs in Nature, etc.), full citations.
  • EVOLUTION.md — version-by-version journey from v0.1 (−34 %, 27 % compliance floor) to v0.13 (−80 %, 87 % compliance floor). Variance-shrinkage centerpiece (66 pp → 13 pp = 5× reduction in cross-agent variance).
  • BENCHMARKS.md — full per-version per-agent raw data for plotting / analysis.

Quick install

```bash

Drop TAUT.md into each agent's global instruction file

cp TAUT.md ~/.claude/CLAUDE.md
cp TAUT.md ~/.codex/AGENTS.md
cp TAUT.md ~/.gemini/GEMINI.md
cp TAUT.md ~/.factory/AGENTS.md
cp TAUT.md ~/.pi/agent/AGENTS.md
cp TAUT.md ~/AGENTS.md

For openclaw + hermes: append after their existing system prompt — see AGENT-LOCATIONS.md

```

Methodology

  • 8 production coding-agent CLIs (Claude Code, Codex, Gemini, Droid, Cursor, Pi, Hermes, OpenClaw)
  • 15-prompt frozen suite covering 13 distinct verbosity-trap categories
  • N=3 trials per (agent, prompt) for trial-stable agents
  • ~3 900 measured agent responses across the v0.1 → v0.13 iteration
  • Cross-agent fair tokenizer: `tiktoken o200k_base`
  • Strict ALL-trials-pass compliance metric (variance-aware)

Credit

Built on the work of caveman by Julius Brussee. TAUT diverges by replacing the caveman persona with a senior-engineer register — see PHILOSOPHY.md §3 for the full design-divergence rationale.

Assets 2
Loading