Auto-Improve: How AI Rewrites Content Until It Hits Your Target Score
Set a target score, set max iterations, and AI rewrites and re-scores automatically until your content meets the bar. Here's how it works and when to use it.
The manual content improvement cycle works but it's slow: submit content, read the feedback, revise, re-submit, read more feedback, revise again. Each cycle takes 5-10 minutes of the writer's time. For content that starts at 55/100 and needs to reach 80, that's 3-4 manual iterations.
Auto-Improve collapses this into a single action. Set a target score (e.g., 85/100), set a maximum number of iterations (1-5), and the AI rewrites → re-scores → repeats automatically until the target is met or iterations are exhausted. You see the full score journey — 62 → 71 → 78 → 85 — and the final version with all improvements applied.
The writer reviews the final output, checks the diff against the original, and decides whether to accept the improvements. The AI does the mechanical revision work. The human makes the final call.
How Auto-Improve Works
The Loop
- Content is scored against your evaluation criteria → initial score (e.g., 62/100)
- AI reads the scored feedback — specific per-criteria issues with quoted text and suggestions
- AI rewrites the content to address the identified issues
- Improved content is re-scored → new score (e.g., 71/100)
- If the new score is below the target → loop back to step 2 with the improved content
- If the new score meets or exceeds the target → stop and present the final version
- If maximum iterations are reached → stop and present the best version achieved
What You See
After Auto-Improve runs, you see:
The score journey:
| Iteration | Overall | Brand Voice | Readability | Accuracy | SEO | CTA |
|---|---|---|---|---|---|---|
| Original | 62 | 70 | 48 | 85 | 60 | 40 |
| Iteration 1 | 71 | 74 | 62 | 84 | 68 | 58 |
| Iteration 2 | 78 | 78 | 72 | 86 | 75 | 68 |
| Iteration 3 | 85 | 82 | 80 | 87 | 82 | 78 |
The diff view: Side-by-side comparison of your original content and the final improved version. Every change is highlighted so you can review what was modified.
Per-iteration changes: What each iteration addressed — "Iteration 1 shortened 6 sentences averaging 32 words to under 20 words. Iteration 2 rewrote the CTA from generic 'Try it free' to article-specific 'Build your first compliance reviewer — free.' Iteration 3 restructured paragraph 4 for better SEO heading hierarchy."
Configuration Options
| Setting | What It Controls | Recommended |
|---|---|---|
| Target score | The score the content must reach to stop iterating | 80-85 for most content; 90+ for compliance |
| Max iterations | Maximum number of improvement cycles | 3-5 (diminishing returns beyond 5) |
| Preserve sections | Content sections that shouldn't be modified | Direct quotes, data tables, legal disclaimers |
When to Use Auto-Improve
High-Volume, Standard Content
Content that follows a predictable pattern and needs to meet consistent quality criteria. Auto-Improve is most effective when the issues are systematic (readability, structure, terminology) rather than creative (voice, narrative, persuasion).
Good fit:
- Product descriptions that need consistent formatting and accuracy
- Support documentation that needs clear, step-by-step structure
- Internal communications that need readability improvement
- Template-based content (email templates, standard reports)
First-Draft Polish
Writers produce a solid first draft but it needs mechanical improvement — shorter sentences, better heading structure, stronger CTAs. Auto-Improve handles these improvements faster than manual revision.
Good fit:
- Blog posts that are well-researched but need readability polish
- Marketing emails that have the right message but weak structure
- Reports that are accurate but dense
Batch Processing
When you have multiple pieces of content that all need improvement to meet the same standard. Run Auto-Improve on each piece rather than manually revising them one by one.
Good fit:
- Updating 20 product pages to meet new brand guidelines
- Improving a batch of blog posts for readability after a style guide change
- Bringing legacy content up to current quality standards
When NOT to Use Auto-Improve
Creative Content
Auto-Improve optimises for criteria scores. Content that relies on creative voice, unexpected phrasing, or deliberate rule-breaking may lose its personality through automated improvement. A thought leadership piece that intentionally uses long, complex sentences for rhetorical effect will have those sentences shortened — technically improving readability but destroying the intended style.
Better approach: Use single-step Improve with AI to get suggestions, then manually select which to apply.
Content Above 85
Diminishing returns are real. Getting from 60 to 80 addresses clear, mechanical issues — sentence length, passive voice, missing elements. Getting from 85 to 95 requires nuance that AI handles less reliably — perfect word choice, subtle tone shifts, strategic emphasis.
Better approach: Use Auto-Improve to reach 80-85, then manual editing for the final polish.
Content Where Accuracy Is Critical
Auto-Improve may rephrase claims in ways that subtly change their meaning. A sentence that says "studies suggest a correlation" might become "research shows a link" — which sounds similar but is a stronger claim than the original.
Better approach: Use Auto-Improve with "Accuracy" as a high-weight criterion and carefully review the diff for any meaning changes. Or use single-step improvement and review each change individually.
Content with Specific Formatting Requirements
Legal documents, regulatory filings, and content with precise formatting requirements (specific clause numbering, mandated section headers) may be altered in ways that break formatting compliance.
Better approach: Mark formatting-critical sections as "preserve" so Auto-Improve doesn't modify them.
The Score Journey: Patterns and Interpretation
Pattern 1: Steady Climb (Ideal)
62 → 71 → 78 → 85
Each iteration addresses the lowest-scoring criteria without degrading others. This pattern indicates well-defined criteria with clear, non-conflicting improvement paths.
Pattern 2: Plateau
62 → 74 → 76 → 77
Big improvement in the first iteration, then diminishing returns. The initial issues were clear and mechanical. The remaining gap requires human judgement or criteria that AI can't systematically improve (originality, strategic messaging).
What to do: Accept the plateaued score if it passes your quality gate. Use manual editing for further improvement.
Pattern 3: Oscillation
62 → 73 → 68 → 75
Score goes up, then down, then up again. This happens when criteria conflict — improving readability (shorter sentences) reduces accuracy scores (oversimplified claims), then improving accuracy adds complexity that reduces readability.
What to do: Review which criteria are conflicting. Either adjust the criteria to reduce conflict, or set priorities in the system prompt ("when readability and accuracy conflict, preserve accuracy").
Pattern 4: Ceiling
62 → 78 → 79 → 79
Quick improvement then hard ceiling. The content has fundamental structural or conceptual issues that iterative rewriting can't fix — it needs a different approach, not better sentences.
What to do: Read the per-criteria feedback for the lowest-scoring areas. If the issues are structural (wrong angle, missing sections, wrong audience), manual rewriting is needed.
Diminishing Returns: The Data
Based on typical improvement patterns across content types:
| Iteration | Average Score Improvement | Cumulative Improvement |
|---|---|---|
| 1 | +8 to +12 points | +8 to +12 |
| 2 | +5 to +8 points | +13 to +20 |
| 3 | +3 to +5 points | +16 to +25 |
| 4 | +1 to +3 points | +17 to +28 |
| 5 | +0 to +2 points | +17 to +30 |
The practical ceiling is around 85-90 for automated improvement. Content starting at 60 can reliably reach 80-85 through Auto-Improve. Getting above 90 requires human editorial judgement.
Recommended max iterations: 3 for most content. 5 for content that starts very low (below 50) or has many systematic issues.
Cost Considerations
Each Auto-Improve iteration uses credits for:
- The improvement rewrite (based on content length)
- The re-score (same cost as a regular review)
For a typical blog post (2,000 words) with 3 iterations:
- 3 improvement rewrites + 3 re-scores = approximately 6x the cost of a single review
Is it worth it? Compare to the alternative: 3 manual revision cycles × 10 minutes each = 30 minutes of writer time. At $50/hour, that's $25 of writer time. Auto-Improve handles the same work for a fraction of the cost and in a fraction of the time.
For high-volume content (50+ pieces/month), the time savings compound significantly.
Tips for Best Results
Set Realistic Targets
A target of 95/100 will use all iterations and likely not reach the target for most content. Set targets based on your quality gate plus a small buffer:
| Quality Gate | Suggested Auto-Improve Target | Rationale |
|---|---|---|
| 70 | 78-80 | Passes gate comfortably; room for minor degradation in final edits |
| 75 | 82-85 | Same logic |
| 85 | 88-90 | Higher gates need smaller buffers — diminishing returns above 85 |
Use High-Weight Criteria Strategically
Auto-Improve prioritises the criteria that contribute most to the overall score. If readability has the highest weight, most improvement effort goes to readability. Make sure your weights reflect your actual priorities.
Review the Diff Every Time
Auto-Improve is not fire-and-forget. Always review the diff view to verify:
- Meaning hasn't changed (especially for claims, data, and technical content)
- Brand voice is preserved (automated rewriting can flatten personality)
- Formatting is intact (especially for structured documents)
- No information was removed that should have been kept
Combine with Knowledge Bases
Auto-Improve produces better results when the reviewer has a knowledge base. Without one, improvements are based on general writing rules. With your brand guide uploaded, improvements align with your specific voice. With compliance rules uploaded, improvements maintain regulatory terminology while improving readability.
Frequently Asked Questions
How long does Auto-Improve take?
Each iteration takes 30-90 seconds depending on content length and model. A 3-iteration Auto-Improve on a 2,000-word blog post typically completes in 2-4 minutes total.
Can I stop Auto-Improve mid-process?
Yes. You can review interim results and stop the process if you're satisfied with the current iteration's score, or if you see the score plateauing.
What if Auto-Improve makes the content worse?
Each iteration's version is saved. If iteration 3 scores lower than iteration 2 (oscillation pattern), you can select the iteration 2 version instead. You're never locked into the final iteration.
Does Auto-Improve change the meaning of my content?
It shouldn't, but it can subtly shift emphasis or strength of claims. Always review the diff view. If accuracy is critical, set it as your highest-weight criterion — the AI will prioritise preserving accuracy over other improvements.
Can I protect specific parts of my content from being changed?
Yes. Mark sections as "preserve" and Auto-Improve won't modify them. Use this for direct quotes, data tables, legal disclaimers, and any content that must remain exactly as written.
Is Auto-Improve better than manual Improve with AI?
For different use cases. Auto-Improve is better for mechanical improvements across multiple criteria when you want to reach a target score automatically. Manual Improve with AI is better when you want to review and approve each round of changes individually, or when the content needs careful, selective improvement.
How many credits does Auto-Improve use?
Approximately 2x credits per iteration (one for the rewrite, one for the re-score). A 3-iteration Auto-Improve uses roughly 6x the credits of a single review. Still significantly cheaper than the equivalent manual revision time.
Should I always use Auto-Improve instead of manual revision?
No. Use Auto-Improve for systematic, criteria-based improvements (readability, structure, terminology). Use manual revision for creative improvements, nuanced tone adjustments, and content where you need full control over every change. Many teams use Auto-Improve to reach 80, then manually polish to 90.
Key Takeaways
- Auto-Improve automates the revision cycle: set a target score, set max iterations, and AI rewrites → re-scores → repeats until the target is met.
- The score journey is visible: see every iteration's score, understand what each cycle addressed, and compare the final version against the original.
- Diminishing returns above 85-90. Auto-Improve reliably gets content from 60 to 80-85. Getting above 90 requires human editorial judgement.
- Best for systematic, mechanical improvements: readability, structure, terminology, completeness. Not ideal for creative polish or nuanced tone.
- Set realistic targets: quality gate + 5-10 points. A target of 95 will exhaust iterations without reaching the goal for most content.
- Always review the diff. Auto-Improve is not fire-and-forget — verify meaning, brand voice, and formatting are preserved.
- 3 iterations is the sweet spot for most content. Beyond 5, you're spending credits for marginal improvement.
- Combine with knowledge bases for context-aware improvements that align with your specific brand, compliance, and industry standards.