Nuwa's output has sources but no testimonial authority
The project asks an agent to model how a public figure reasons. It collects authored work, long conversations, short-form expression, outside criticism, decisions, and a timeline, then extracts recurring mental models and decision rules. The finished skill is supposed to label direct statements, outside reports, and inference separately. That is a better premise than feeding a famous name to a chatbot and accepting whatever voice appears, though the result remains a model assembled from selected public evidence.
Language is the first practical detail. The root README and 38,280-byte SKILL.md are Chinese, while README_EN.md gives English readers a substantial guide. Our commit fe03746 checkout had 159 files, about 1,812 lines counted as source, and occupied 35.3 MB. Much of the repository is example research and media rather than executable Python. That helps readers inspect finished profiles, but it also makes installing the root skill heavier than its runtime instructions alone would require.
Six research lanes make provenance visible
Nuwa splits research across writing, conversations, expression, external views, decisions, and chronology. Each agent writes into a named file under references/research/, and the workflow prefers books, original interviews, and observed decisions over summaries. It tells researchers to preserve contradictions and mark confidence. For Chinese figures, it rejects Zhihu, WeChat public accounts, and Baidu reference pages, while favoring original video, podcasts, and named media outlets. Those source rules are opinionated enough to audit.
The workflow also distinguishes fast, standard, and deep runs before work begins. Its own instructions warn that a full multi-agent job can cost tens of dollars on top models. That cost claim belongs to the project, not our lab. Our sandbox had 3 CPUs and 8 GB, but it measured repository setup rather than launching the 6-agent research process. Buyers should expect model calls, browsing time, source downloads, and manual review to dominate the five-second build step.
What happened when we ran it
Our sandbox installed 35 Python packages in 16 seconds and used 37 MB for the installed environment. The detected build step succeeded in 5 seconds. Pip-audit found 0 known vulnerabilities. Those results show that commit fe03746 can complete the harness's setup and build phases in a fresh Python 3.12 Bookworm image without secrets. They do not measure research quality, generated-skill accuracy, or the cost of running subagents.
There was no test script or target, so the lab skipped tests. The checkout had no tests directory and no Dockerfile, while one CI workflow checks community-index pull requests. That workflow deliberately checks out the repository's own main-branch script under pull_request_target instead of executing contributor code. It is a thoughtful control for the contribution path, but it does not provide regression evidence for extraction, source handling, path selection, or the generated perspective's answers.
The fidelity scorecard tests resemblance through another model
Generated examples include a scorecard that grades stance consistency, recognizable style, honesty at the edge of available evidence, source transparency, and structural completeness. The answering agent and scoring agent must be separate, and prompts should include questions with known public answers plus one outside the person's documented scope. This can catch a profile that turns into a generic assistant or states an invented view with false confidence. Keeping the result in FIDELITY.md also makes later comparisons possible.
The score is still an LLM judgment, not identity verification. Nuwa's README reports 15 official skills at grade A, while our lab ran 0 fidelity cases because no test target existed. A profile can match familiar positions and speaking habits while omitting a contrary source or projecting an old stance onto a new event. For consequential use, read the citations behind the answer and ask for uncertainty. Never convert a generated sentence into a quote attributed to the modeled person.
Broad runtime claims meet Claude-specific paths
Installation is advertised through npx skills add alchaincyf/nuwa-skill, manual cloning, or pasting the Markdown into an agent. The README names Claude Code, Codex, Cursor, OpenClaw, Hermes, Gemini CLI, and other skills-aware systems. Inside the workflow, however, Nuwa creates new profiles under .claude/skills/ and scans that same directory for existing perspectives and helper skills. A non-Claude runtime may load Nuwa successfully yet write the result where that runtime never discovers it.
Our 5-second build does not resolve that portability gap. Adjust the target directory before starting a costly run, then verify that the chosen agent can spawn parallel researchers and save every source file. GitHub showed 2 open issues and 4 open pull requests, with the last push on August 25, 2026. Issue 74 specifically asks the project to separate runtime files from roughly the media-heavy root. There is no tagged release, so pin fe03746 or another reviewed commit instead of following main silently.