Tools: InkSafe 2.0 what I rebuilt after the first launch and why the AI is completely different now

After my first dev.to post about InkSafe got some readers I did what I probably should have done before launching — I sat down and stress tested the AI properly. What I found was uncomfortable. The original version was good at catching obvious red flags but it was missing subtle issues, it wasn't explaining the real world consequences clearly enough, and it had a JSON parsing bug that occasionally made results look like someone had dropped a keyboard into the output.
So I rebuilt it.
Here's everything that changed in InkSafe 2.0.
The AI model The original version ran on a lighter, faster AI model that sometimes missed nuance in complex legal language. 2.0 runs on a significantly more powerful model. The difference in output quality is noticeable — better at understanding context, better at catching subtle language, better at explaining what a clause actually means for the person signing it. The analysis prompt — rebuilt from scratch
This was the biggest change. The original prompt was good but vague in places. The new version has a 30-point checklist that the AI works through on every single contract: Termination rights and notice periods. Payment forfeiture on termination. Subjective payment approval. Payment timeline. IP assignment scope. IP transfer timing. Template and workflow ownership. Non-compete scope and duration. Non-solicitation. Exclusivity. Warranty period. Liability cap imbalance. Consequential damages exposure. Indemnification balance. Penalty clause proportionality. Scope modification rights. Revision limits. Portfolio and publicity rights. Mandatory social media obligations. Storage and archiving obligations. Auto-renewal notice periods. Unilateral contract modification. Arbitration location and costs. Governing law and legal cost allocation. Name and likeness usage. Backup or substitute restrictions. Cancellation penalty proportionality. Equipment and delivery liability. Confidentiality scope and duration. Assignment rights.
Every contract gets checked against all 30 points. No shortcuts.
Compound risk detection — the most important upgrade
This is the change I'm most proud of.
The old version flagged individual clauses. A non-compete here. A payment issue there. What it missed was how clauses interact with each other to create compounded risk.
Here's a real example of what I mean.
A contract might have two clauses that each look manageable on their own:
Clause 4: Client is entitled to unlimited revisions until fully satisfied.
Clause 6: Final payment due upon Client's written acceptance of all deliverables. Neither clause alone is catastrophic. But combined they create a situation where a client can demand infinite rounds of revisions indefinitely while withholding 100% of the final payment — because they haven't accepted the deliverables yet. You could work for months and legally receive nothing.
InkSafe 2.0 now identifies these combinations explicitly and explains the compounded danger, not just the individual clauses.
Other dangerous combinations the AI now catches:
Termination without notice plus payment forfeiture — client can take your work and pay nothing. IP assignment of templates plus exclusivity — you lose your tools and cannot work elsewhere simultaneously.
Penalty clauses plus no equivalent client penalty — you bear all the financial risk, they bear none.
Mandatory social media posting plus confidentiality — contradictory obligations that put you in breach no matter what you do.
What InkSafe 2.0 now scans
The original version was built primarily around developer and designer contracts. The new version handles everything a freelancer gets asked to sign: Developer agreements, design contracts, photography SOWs, video editing retainers, consulting MSAs, copywriting contracts, brand strategy agreements, NDAs, subcontractor agreements, and general independent contractor agreements. The AI also adjusts its analysis based on what type of freelancer the contract is for. A photography contract gets scrutinized for RAW file demands and equipment liability. A developer contract gets scrutinized for pre-existing code library grabs and open source warranty traps. A design contract gets scrutinized for template ownership and unlimited revision traps. The bug fix nobody noticed but everyone experienced
The JSON parsing bug. If you used InkSafe before the fix you might have seen results that ended with a wall of characters like }b}nalysis}e}s}result} after the summary.
That's fixed. Clean results every time now.
What the scoring looks like in practice
Since rebuilding the prompt I've run the AI through seven different contract types. Here's the range:
A 15-clause developer contract loaded with traps — 5/100, 9 issues caught. A photography contract with one-sided cancellation penalties and unlimited consequential damages — 12/100, 7 issues caught.
A video editing retainer with template ownership grabs and impossible turnaround penalties — 10/100, 7 issues caught.
A consulting SOW with exclusivity that contradicted independent contractor status — 15/100, 7 issues caught.
A balanced, well-written freelance agreement — 88/100, 3 minor improvements suggested. The scoring works in both directions. A genuinely fair contract gets a good score. The AI isn't alarmist. What's next The core product is solid now. The next focus is getting real users and real feedback on contracts I haven't written myself.
If you're a freelancer with a contract sitting in your inbox right now I'd genuinely love you to run it through InkSafe and tell me what the AI gets wrong, misses, or explains badly. That feedback is worth more to me than any feature I could build right now.
Free to try at inksafe.ai — one free scan, no credit card needed. As always — AI analysis for informational purposes only, not a substitute for qualified legal advice. Templates let you quickly answer FAQs or store snippets for re-use. Are you sure you want to hide this comment? It will become hidden in your post, but will still be visible via the comment's permalink. Hide child comments as well For further actions, you may consider blocking this person and/or reporting abuse