Skills / wiki-generator

Safety — wiki-generator

Read this before running the pipeline unattended or adapting it to your own material.

Data handling

The workflow sends source-derived text and generated images through external products: ChatGPT (summaries, images), Google Drive (all shared state), and your agent's model provider (integration and review). Anything you feed into the pipeline is visible to those providers.
The generated wikis are private. This pipeline is a personal reading companion, not a publishing pipeline. If you publish output, you own checking the rights to the source material first.
Source books must be legally obtained (purchased EPUBs, your own audiobooks). Do not feed pirated or leaked material into it.
Do not reuse this exact pattern for confidential client material without reviewing provider retention, permissions and access controls first. The failure knowledge in these files comes from recreational content; the same transparency does not exist for regulated data.

The integration cron runs hourly with filesystem access, git push rights, and Drive write access. Scope its credentials to what it needs: a Drive account that can only see the bridge folders, git remotes that only hold the wikis.
The review command in wiki-builder.md uses --dangerously-skip-permissions (with script -qc for PTY allocation). That flag is an operator choice with consequences, not a default recommendation: it disables the CLI's permission prompts for that call. Here it is accepted for a short, read-only review call on the operator's own machine, in a workflow where the prompt is fixed and the inputs are only local image files. If you copy this pattern, first try running without the flag and only add it if a permission rule genuinely cannot be made to work.
prompts.md is volatile by design; never treat it as the only copy of anything. The queue files hold the durable work state.

Transcripts are kept uncorrected as a faithful record of what the audio contained; corrections happen downstream in the wiki pages. Do not "fix" a transcript in place.
Progressive building is a spoiler-safety property, not just an aesthetic one: chapter pages are generated with only the knowledge of chapters processed so far. Breaking it (e.g. by letting the model use external knowledge of the book) leaks later-chapter information into early pages.