AI agents often produce grammatically perfect but spiritually hollow text. The culprit isn't syntax; it's an unearned earnestness and a predictable rhythm that screams 'synthetic.' To combat this, AI engineer Yadlapalli Avinash Ricky has released 'writ,' an open-source self-audit skill designed to strip these specific habits from machine-generated drafts. The tool codifies findings from Wikipedia editors who cataloged thousands of AI submissions into a guide called 'Signs of AI Writing,' turning those observations into an actionable workflow for developers.

The Anatomy of Synthetic Prose

The core issue with LLM output is 'over-narrated significance.' Models trained via reinforcement learning from human feedback (RLHF) tend to sound like anxious corporate press releases. Instead of stating facts, they add participatory tails like 'highlighting the transformative shift' or 'underscoring the critical need.' These dangling participial clauses manufacture weight where none exists. If you delete the clause, the factual value remains unchanged, proving the text was narrating its own importance rather than delivering information.

Six Patterns That Give It Away

The 'writ' skill targets six specific tells. First, it kills dangling clauses ending in '-ing' phrases like 'marking a pivotal moment.' Second, it bans stock significance vocabulary including 'delve,' 'robust,' 'tapestry,' and 'seamless.' Third, it removes false contrasts such as 'It's not just a databaseβ€”it's the backbone.' Fourth, it limits em dash overuse, which models use to splice thoughts unnaturally. Fifth, it eliminates defensive hedging around settled facts. Finally, it breaks rhythmic uniformity, forcing variation in sentence length to avoid the metronomic cadence typical of AI prose.

Integrating the Audit Into Your Workflow

Unlike generators that create text from scratch, 'writ' acts as an editorial second pass. It ingests existing drafts, scans them against the codified heuristics, and rewrites passages to preserve technical intent while removing synthetic markers. For example, a draft describing Redis might transform from 'seamlessly storing frequently accessed records... underscoring the critical need' to 'Redis stores hot keys in memory to keep read latencies under two milliseconds.' The latter is shorter, specific, and free of posturing.

How to Deploy 'writ'

Developers can integrate 'writ' in three ways. Agentic coding tools like Cursor, Antigravity, and Claude Code can invoke the skill via a slash command like '/writ' to review drafts. For manual workflows, users can paste the SKILL.md file from the GitHub repository into their system prompts. Additionally, teams can add 'writ' as an automated review step in CI documentation pipelines to flag PR descriptions or markdown docs that lean too heavily on marketing adjectives before they merge.

Key Takeaways

  • Facts carry their own weight; specific benchmarks convince readers better than words like 'groundbreaking.'
  • Read sentences aloud to catch uniform rhythm that eyes might miss.
  • Audit after drafting rather than self-censoring during the initial write.
  • Replace abstract descriptors with precise nouns and verbs, such as swapping 'robust pipeline' for 'pipeline handling 50,000 events per second.'

The Bottom Line

AI agents are excellent at syntax but terrible at soul. 'writ' doesn't make the AI smarter; it just stops it from lying about how important it is. For technical writers, this is the first practical step toward prose that doesn't immediately trigger a reader's 'delete' reflex.