A controlled-language variation of ASD-STE100 for software documentation — rules for writing comments, commit messages, docs, error strings, and config in plain, unambiguous English that LLMs and humans both follow reliably.
| Stage | Dir | What it produces |
|---|---|---|
| A Extraction | ste-code/extracted/ |
raw text pulled from the spec pages |
| B Refinement | ste-code/refined/ |
cleaned, section-aware markdown |
| C Grouping | ste-code/grouped/ |
deterministic regrouping into the 9 rule sections (no LLM) |
| D Adaptation | ste-code/adapted/ |
ASD-STE100 rules rewritten for the code domain |
| E Extensions | ste-code/final/extensions/ |
code-domain vocabulary + anti-pattern gap fillers |
| F Artifacts | ste-code/artifacts/ |
LLM-distilled, deployable tiers (this is the deliverable) |
ste-code/final/ is the enriched consolidated standard that Stage F is built
from. ste-code/spec/ holds the source spec pages.
ste-code/artifacts/ ships 8 tiers (level-2 … level5/), each a
directory of small sub-documents plus a system-prompt.txt (the whole tier in
one file). llms.txt and llms-full.txt are the consolidated single-file
forms. See ste-code/artifacts/README.md.
.agents/benchmark/ runs the 59-test / 14-category suite against any tier:
python3 .agents/benchmark/benchmark-levels.py --levels -2,-1,0,1,2,3,4,5
python3 .agents/benchmark/orchestrator-control.py # plain-assistant baselineAll pipeline tooling lives in .agents/tools/ and is agent-agnostic —
backends (Hermes, Claude, Codex) are configured in .agents/config/agents.yaml.
ste-code/templates/ holds the externalized prompt templates consumed by the
trajectory system (.agents/tools/trajectory/). They are rendered at run time,
not hand-edited per run.