aus-gov-style-rules

The Australian Government Style Manual (AGSM), compiled as portable, AI-native rules.

Not an app. Just the rules — in formats you can paste into a system prompt, feed to a rule engine, or reference from any tool stack.

---

Quickstart

Paste into any LLM — copy prompts/quick.md into your system prompt. Done.

Use the rule engine — load schema/ruleset.json and run it against text. See examples/pipeline-usage.md.

---

What's in here

aus-gov-style-rules/
│
├── prompts/
│   ├── quick.md              ← top ~20 mechanical rules, minimal tokens — paste anywhere
│   ├── full.md               ← complete all-category AGSM system prompt, ~15KB
│   └── by-category/
│       ├── grammar-punctuation.md
│       ├── numbers-dates.md
│       ├── capitalisation.md
│       ├── plain-english.md
│       ├── inclusive-language.md
│       ├── referencing.md
│       └── structuring.md
│
├── schema/
│   ├── ruleset.json          ← 270 machine-runnable rules (cookbook rule engine format)
│   └── extraction-schema.json
│
├── rules/                    ← 251 source rules in YAML, one file per category
│   ├── grammar-punctuation.yaml
│   ├── numbers-dates.yaml
│   ├── capitalisation.yaml
│   ├── plain-english.yaml
│   ├── inclusive-language.yaml
│   ├── referencing.yaml
│   └── structuring.yaml
│
└── examples/
    ├── before-after.md       ← 40+ worked violation + correction examples
    └── pipeline-usage.md     ← how to run the rule engine (Python code)

---

Three ways to use it

1. System prompt drop-in (zero setup)

Copy prompts/quick.md into the system prompt of any LLM. It covers the top ~20 mechanical AGSM rules — dates, time, numbers, percentages, currency, capitalisation, plain English, inclusive language — in the most compact form.

For a complete drop-in (all 7 categories), use prompts/full.md.

For a targeted prompt (one topic), use a file from prompts/by-category/.

Workflow:
  1. Open prompts/quick.md
  2. Copy the contents
  3. Paste into your LLM system prompt
  4. Done — the LLM now enforces AGSM rules

2. Rule engine (pipeline integration)

schema/ruleset.json contains 270 compiled rules in a structured format suitable for automated checking.

Rule types:

Type	Count	How it runs
regex	104	Pattern match — no model needed, instant
llm-deterministic	92	Yes/no judge question — fast model (e.g. Haiku)
llm-assertion	74	Full evaluation prompt — capable model (e.g. Sonnet)
Total	270

See examples/pipeline-usage.md for complete Python code.

Quick example:

import json, re

with open("schema/ruleset.json") as f:
    ruleset = json.load(f)

# Run regex rules instantly — no model
text = "The unemployment rate fell by 0.3 percent last quarter."
violations = []
for rule in ruleset["rules"]:
    if rule["check_type"] == "regex":
        if re.search(rule["when"]["value"], text, re.MULTILINE):
            violations.append(rule["id"])

print(violations)
# ['percentage-per-cent-two-words']

3. YAML source (build your own tooling)

The rules/ directory contains 251 source rules in human-readable YAML, one file per AGSM category. Each rule has:

• id — stable kebab-case identifier
• type — deterministic or llm-assertion
• title, description — in plain English, not AGSM prose
• test.pass / test.fail — example text
• severity — error, warning, or suggestion
• source — AGSM page URL
• agsm_source_date — when the rule was compiled

- id: percentage-per-cent-two-words
  type: deterministic
  category: numbers-dates
  title: Write 'per cent' as two words in Australian English
  description: >
    The spelled-out form in Australian English is always two words: per
    cent. The single-word form "percent" is not standard Australian spelling.
  test:
    pass:
      - "Fifty-five per cent of the council's revenue came from rates."
    fail:
      - "Fifty-five percent of the council's revenue came from rates."
  severity: error
  source: https://www.stylemanual.gov.au/grammar-punctuation-and-conventions/numbers-and-measurements/percentages
  agsm_source_date: "2026-09-22"

---

Rule coverage

Category	YAML rules	Ruleset entries
Grammar + punctuation	48	61
Numbers + dates	43	41
Capitalisation	30	29
Plain English	30	30
Inclusive language	38	39
Referencing	28	28
Structuring	34	42
Total	251	270

Ruleset entries exceed YAML source rules because the llm-deterministic and llm-assertion splits produce separate compiled entries from the same source.

---

Worked examples

examples/before-after.md has 40+ worked before/after examples across all categories. A sample:

Date format

❌ The consultation period ends December 31, 2026.
✅ The consultation period ends 31 December 2026.
   Rule: date-format-day-month-year

Nominalisation

❌ Please make an application before the closing date.
✅ Please apply before the closing date.
   Rule: avoid-nominalisation

Disability language

❌ Disabled people can access support through this service.
✅ People with disability can access support through this service.
   Rule: disability-person-first-default

---

Related: Octavius

Octavius (IP Australia) is a full Streamlit + spaCy application for AGSM linting with deterministic regex/NLP rules.

This repo is a different thing, not a competitor:

	Octavius	aus-gov-style-rules
Type	App (Streamlit + spaCy)	Portable data (no app)
Rules	Deterministic (regex/NLP)	Deterministic + LLM-assertion
Output	Running linter	System prompts, YAML, JSON
Use case	Run it locally	Drop into any tool stack
LLM-aware	No	Yes — evaluation prompts included

If you want a linter to run locally, use Octavius. If you want rules to paste into Claude or drop into a pipeline, use this.

---

How rules are compiled

1. Claude reads each AGSM category page from stylemanual.gov.au
2. Rules are re-expressed in original language — the model extracts meaning and writes rules in its own words; AGSM prose is never copied verbatim
3. Each rule cites its AGSM source URL and carries an agsm_source_date
4. Deterministic rules are compiled into the rule engine format with regex patterns and LLM judge questions
5. LLM-assertion rules carry evaluation prompts — self-contained instructions for a capable model to judge violations

The AGSM is a living document — rules here reflect the site as of the date in AGSM-VERSION.md. To update, re-extract the relevant category.

---

Copyright and licensing

The AGSM is © Commonwealth of Australia 2020, published under standard Crown copyright with no Creative Commons licence.

Rules in this repository are not verbatim reproductions of AGSM prose. They are re-expressed in original language compiled by AI, citing source URLs. Rules themselves (principles, facts) are not copyrightable — only specific expression is.

This repository is dual-licensed:

• MIT — tooling, scripts, schema definitions
• CC-BY 4.0 — compiled rules content (rules/, prompts/, schema/ruleset.json)

Attribution: bitsloppy/aus-gov-style-rules

See AGSM-VERSION.md for full copyright notes.

---

Contributing

Rules extracted using Claude with a consistent schema. To contribute:

• Keep to the re-expression approach — no verbatim AGSM prose
• Follow the YAML schema in schema/extraction-schema.json
• Include source URL and agsm_source_date on every rule
• Add pass/fail examples that unambiguously test the rule

Open an issue if you spot an error or want to add a rule category.