Every few years a tool gets good enough that "can we build this?" stops being the interesting question and "should we, and who signs off?" takes over. That is exactly where AI coding agents now sit with retro-game reimplementation. Anthropic's recent adjustment to Claude Code's system prompt — steering the agent away from re-creating existing video games while still allowing original game development — reads less like a product decision and more like a signal: model vendors are now willing to bake legal-risk heuristics directly into the agent's instructions.
For engineering teams, the interesting question is not whether you agree with that stance. It is whether your own workflow can answer a simple question under pressure: what did the agent build, from what source, and who approved it?
Below is a checklist you can run this week. Every item has an explicit pass condition and a fail condition. A fail is not a moral judgment — it is a gap that will cost you days the moment a rights holder, a client, or a platform reviewer asks.
What changed, and why it lands on your team
The tip matters beyond gaming. Retro ports, decompilations and recompilations are just the most visible collision between generative tooling and decades-old intellectual property. The same pattern shows up in UI clones, API re-implementations, asset packs, and "make me a version of [popular product]" prompts that arrive from product managers rather than engineers.
When a vendor narrows what its agent will do, it is effectively telling you that the boundary is now part of the toolchain contract. That cuts both ways: it reduces some risk automatically, and it removes the excuse that nobody told you. If your team's only control is "the model probably won't," you have no control at all.

Figure: Editorial context image from the source page (preview only; rights remain with the publisher) · Credit / further reading: www.theverge.com
The checklist below is deliberately mechanical. You should be able to hand it to a tech lead on a Monday and have a yes/no answer for every line by Friday.
The checklist: six gates, pass or fail
Gate 1 — Scope intake
- Source declaration. Every agent-assisted task starts with a one-line statement of what is being built and what it is derived from.
- Pass: "Original arcade-style shooter; no reference title; no extracted assets."
- Fail: "Make something like [named commercial game]" with no further annotation.
- Named-reference flagging. Any prompt that names a commercial title, studio, or specific game mechanic triggers a review before code generation, not after.
- Pass: The flag exists in your ticket template and blocks the task from moving to "in progress."
- Fail: The reference is discovered in a commit message three weeks later.
- Clean-room intent. Teams building in a crowded genre document that they are not using disassembled binaries, leaked source, or extracted art as inputs.
- Pass: A signed statement in the repo's
CONTRIBUTINGordocs/legalfolder. - Fail: "We'll figure out the legal side at launch."
- Pass: A signed statement in the repo's
Gate 2 — Agent configuration
- System-prompt parity. Your agent configuration reflects current vendor guidance, not a snapshot from six months ago.
- Pass: A named owner reviews agent instructions on a fixed cadence and records the review date.
- Fail: Nobody knows who last edited the system prompt.
- Repository instruction files. Project-level instruction files (the ones agents read automatically) include an explicit no-cloning clause.
- Pass: The clause is present in the repo, versioned, and visible in diffs.
- Fail: The rule lives in a Slack thread.
- Guardrail testing. You periodically probe whether the agent actually refuses the thing you told it to refuse.
- Pass: A small set of test prompts is run each quarter and results are logged.
- Fail: You assume the guardrail works because the vendor blog said so.
Gate 3 — Provenance of code and assets
- Dependency legibility. Every third-party library, shader, font and sprite in the build has a known origin and license.
- Pass: A machine-readable bill of materials is generated on every build.
- Fail: Provenance lives in one engineer's head.
- Generated-asset labeling. Assets produced by generative models are tagged at the file level, not just in a wiki.
- Pass: Metadata or a sidecar manifest records the generating tool and date.
- Fail: You cannot distinguish generated art from commissioned art.
- Similarity screening. Before shipping, new code and assets are compared against known references.
- Pass: Automated similarity checks run in CI, with human review of flagged items.
- Fail: "It looked different enough."

Figure: HEXOPS original illustration inspired by the topic (not a copy of the source article imagery) · Credit / further reading: www.theverge.com
Gate 4 — Human review gates
- Agent output is never merge-ready by default. AI-authored changes require a named human approver.
- Pass: Branch protection enforces review by someone other than the prompt author.
- Fail: The agent commits directly to a shared branch.
- Scoped diff review. Reviewers check what the agent chose to reuse, not just whether tests pass.
- Pass: The PR template asks: "Does this reproduce any third-party behaviour or expression?"
- Fail: Review means green CI.
- Feature-level sign-off. Gameplay systems that mirror a specific title require a product and legal acknowledgement.
- Pass: A recorded approval with a date and a named person.
- Fail: Approval is implicit because nobody objected.
Gate 5 — Records and audit trail
- Prompt and output logging. You can reconstruct which prompts produced which commits.
- Pass: Logs are retained per your data policy and searchable by task ID.
- Fail: Conversations vanish when the session closes.
- Decision record. Judgement calls about closeness-to-reference are written down.
- Pass: A short decision note exists for each flagged project.
- Fail: The reasoning is only in someone's memory.
- Retention discipline. Logs containing proprietary or user data follow the same retention rules as production data.
- Pass: Retention is defined and enforced by policy.
- Fail: Logs accumulate indefinitely because deleting them is nobody's job.
Gate 6 — Escalation and kill switch
- Defined trigger. Your team knows what constitutes a stop-the-line event (a takedown notice, a platform rejection, a credible similarity claim).
- Pass: The trigger is written down and understood by on-call.
- Fail: Escalation happens ad hoc, on social media first.
- Reversible rollout. Generated content ships behind a feature flag or staged release.
- Pass: You can pull the affected build within hours.
- Fail: Removal requires a full re-release cycle.
- Vendor-change watch. Someone owns tracking model-policy updates and translating them into repo-level rules.
- Pass: A named owner and a standing agenda item.
- Fail: Policy changes reach you via a news article.
Key takeaways
- Vendor guardrails reduce risk but transfer responsibility: once a tool documents its limits, "we didn't know" stops being a defence.
- The cheapest control is intake — a single line declaring what a task is derived from catches most problems before a token is spent.
- Provenance beats intent. If you cannot list the origin of every asset and dependency in a build, you cannot answer a challenge.
- Human review must be scoped to reuse, not just correctness. Green tests say nothing about originality.
- Audit trails are an engineering feature, not a legal afterthought. Design logging and retention the same way you design backups.
Next steps
- This week: add a source-declaration field to your task template and a no-cloning clause to repo-level agent instructions.
- Two weeks: inventory one active project end to end — dependencies, assets, generated files — and mark every item without a known origin as a fail.
- One month: run a guardrail probe against your configured agent, log the results, and assign a named owner for vendor policy changes.
- Ongoing: put similarity screening into CI and add the reuse question to your PR template.
Sources
- Anthropic clamps down on Claude's AI video game clones — The Verge AI



