Agent Rules

Edit prose so it reads once. Report what changed and why.

The full ruleset is in references/rules.md. Read it before editing. Part 1 covers writing. Part 2 covers working behaviour and applies only when the text you are reviewing is an agent's own report of work it did.

Two modes

Edit is the default. Rewrite the text to follow the rules, then say what you changed.

Check runs when the user asks whether something breaks the rules, or says check, audit, score, or "is this slop". Report violations. Do not rewrite.

Workflow

1. Score the text first, so you have a number rather than an impression:

   python3 scripts/score.py --text draft.md
   

   It reports word count, average sentence length, the share of sentences over 20 words, a Flesch score, and every violation with its line number.

2. Read references/rules.md.

3. In check mode, report the violations and stop.

4. In edit mode, fix them in this order, because the early ones remove the most text:

   • Cut preamble. The first sentence must answer the question.
   • Cut the closing pleasantry and any unrequested offer to help further.
   • Cut sentences that only signal effort or state something nobody would disagree with.
   • Replace banned words with the plain word that was meant. Keep the word when it is the precise technical term.
   • Split every sentence over 20 words. One instruction per sentence.
   • Remove decorative em dashes. Use a comma, a period, or parentheses.
   • Rewrite the named sentence patterns: binary contrast, throat-clearing, faux insight, rhetorical setup, colon reveal, fake-profound ending, summary-recap ending.
   • Convert a real comparison across two or more dimensions into a table. Leave narrative as prose.

5. Score again and show both numbers.

6. Say what you changed. Name the rule behind each edit rather than listing every word you touched.

What not to do

Do not flatten meaning to hit a word count. A shorter draft that dropped a concrete number, a file path, or a caveat is a worse draft. If following a rule would remove real information, keep the information and say which rule you broke.

Do not rewrite quoted material, code, log output, or someone else's words inside the text.

Do not change the author's voice. Bluntness, humour and strong opinions stay. The target is plain, not neutral.

Do not report a violation the checker found if it is a false positive, such as a banned word used as a precise technical term, or a pattern name quoted inside backticks. Say why you are dismissing it.

Do not relabel what the checker found. Use its own category names, so a reader can map your report back to the tool. "Great question!" is a preamble opener, not a banned word.

Always-on alternative

This skill loads only when invoked. Response style needs to apply to every turn, so for that use the global install instead, which writes the same ruleset into each agent's memory file:

git clone https://github.com/tornikegomareli/agent-rules.git ~/.agent-rules
cd ~/.agent-rules && ./install.sh

The two work together. The global rules govern how the agent writes. This skill cleans text that already exists, including text a person wrote.