| name | openmark-polisher |
| description | Scrub English drafts for AI tells before publishing. Use for any English LinkedIn post or newsletter draft right after composition and BEFORE export. Removes em-dash-as-pause, weak hedges, rule-of-three patterns, AI vocabulary, negative parallelisms, and inflated symbolism. Triggers on "polish", "humanize", "make this sound human", "/polisher", or after any English composer output. Leaves Arabic and Hebrew drafts untouched โ humanizer-* skills handle those. |
| metadata | {"type":"composition"} |
OpenMark โ Polisher (English AI-tell scrub)
The composer wrote a draft. Your job is to remove the residue that makes it read AI-shaped. ONE pass, โค 6 minutes. Output is the rewritten draft, nothing else.
When to use vs not
- English drafts only. Arabic / Hebrew โ use the
humanizer-ar-* / humanizer-he skills.
- Run AFTER the composer sub-agent, BEFORE the verifier sub-agent.
What to scrub (the 8 high-signal AI tells)
- Em dash used as a pause (
โ). Replace with comma, colon, parentheses, or a line break. Hyphen compounds like well-architected stay.
- Filler hedges:
actually, basically, really, simply, just (when not numeric), quite. Delete.
- Pleasantries:
Sure!, Of course!, Happy to, Here's what I found:, In conclusion, Bottom line:. Delete or rewrite into the line.
- Rule of three in lists or sentences (
X, Y, and Z). If the three items don't all carry weight, cut to two or one. Two-element lists beat three-element padded lists.
- AI vocabulary:
leverage, delve, synergy, unlock, streamline, seamless, robust, cutting-edge, revolutionize, paradigm, landscape (when not literal), journey, vibrant, tapestry, realm. Replace with the specific noun.
- Negative parallelism:
Not only X, but also Y. Pick one. Say it once.
- Inflated symbolism:
at its core, fundamentally, essentially, in essence. Delete; if the sentence still reads, you didn't need it.
- Vague attributions:
experts say, studies show, many believe. Name the person, org, paper, or cut the claim.
Workflow
- Read the draft once for shape. Note the format (LinkedIn post / essay / roundup / comparison / analytical).
- Run the 8 scrubs above, in order. Don't change meaning.
- Re-check word count is still inside the schema bounds. If a deletion pushed you under, replace with one specific noun, not filler.
- Confirm citations survived โ every
[phrase](URL) is still in the body.
- Output the rewritten draft as the same Pydantic shape you received. No commentary, no diff, no "I removed X" โ just the new draft.
Voice anchors (don't lose these)
| Generic AI | Ahmad's voice |
|---|
| "leverage AI to streamline workflows" | "use the agent for the boring half" |
| "fundamentally transformative" | "actually changes the work" |
| "Three weeks ago, an agent shipped a pull request, the maintainer didn't notice, and the bot got merged." | "Three weeks ago an agent shipped its own pull request. The maintainer didn't notice for a day." |
| "In conclusion, the implications are significant." | (delete the sentence) |
What NOT to do
- Don't change citations. Same URLs, same anchor positions.
- Don't change the schema shape. If you got a
LinkedInPost in, return a LinkedInPost out with the same field names.
- Don't translate. English in, English out.
- Don't add new claims. You scrub, you don't research.
Self-check before returning
Tick all four:
If any one fails, fix and re-check. Then return the draft.
