# Prompt: create a handover file for an abandoned project (v0.3) Give this text to your agent (Claude Code, Codex, Cursor …) while it works inside the project folder. The template `HANDOVER.md` must be next to it or pasted below. --- You are creating a handover file `HANDOVER.md` in the root of this project. A stranger – human or AI agent – should then understand within minutes what the project is, how far along it is, and where to continue. Follow the sections and header of the template. Procedure: 1. **Explore first, ask second.** Read README, docs, plans, commit history (if present), package files, tests and code. Run build and tests if that is safe, and record the result with the date. Do not change source code; build output such as `node_modules/` or `dist/` is fine. Without git history: derive `last_activity` from file dates or docs and mark it `# assumed`. 2. **Separate verified from assumed.** What you saw in code or a test run is fact. Everything else: `(assumed)` in the text, `# assumed` in the header. Never invent features. 3. **Expose contradictions.** If docs and code disagree, the code wins. Name every contradiction concretely in section 10. 4. **Ask the owner only what you cannot know** – in a single round, at most 8 questions. Typical: Why did you stop? How much time/money went in? AI share? License? Offer? Source? If the owner does not answer or you run non-interactively, write `OPEN: ask owner` and list the question under "Open owner questions" – that is the normal non-interactive path. 5. **Scan for secrets and private data – this step is mandatory.** a. Run the scanner: `python3 -I /tools/scan.py . --history --json /scan.json` (falls back to your own search if Python is unavailable). It checks current files and every line ever committed. b. Judge every finding: **real** (must go), **test/example data** (state why you think so), or **false positive**. c. Look for what a pattern scanner cannot see: real names of people or customers, internal hostnames, server addresses, school/company internals, personal photos, screenshots with data, database dumps, config files that are only "example" by name. d. Put the result in section 9 as a table and give the **Publish check** verdict: `BLOCK` (real secret or personal data present – do not publish until removed and keys are rotated), `WARN` (only low-risk items like own server IPs or home paths), or `OK`. A secret found only in git history still means BLOCK: deleting the file does not remove it – the history must be rewritten or the repo published without history, and the key must be rotated. e. Never copy a matched value anywhere – not into files, not into chat, not into your own shell output. If the goal of the project is unknown, measure `completion_percent` against what the code clearly aims at and mark it `# assumed`. f. **Alert the owner immediately** in your final message for every BLOCK item. Never write a value into any file, only type and location. `secrets_removed` stays `null`/`false` until the owner has cleaned up. 6. **Estimate honestly.** `completion_percent` is measured against the project goal, not the last plan step. Hours are always hands-on hours of an experienced human – AI runtimes from logs are not a measure. 7. **Hunt for pitfalls in the code.** TODOs, stubs/placeholders, dead code, cache or version switches that must be bumped on changes, missing auth on services that cause costs. Put them in section 8. 8. **Self-contained.** The file must make sense without the rest of the docs: carry the essentials of long plans/journals into it instead of only linking to them. 9. **Concise.** Aim for 150–400 lines for real projects; tiny projects may be much shorter – never pad. Finished-but-unused projects count too: then section 6 says "finished, not used" and section 7 lists extensions. Tables and lists over prose. Condense long docs and refer to them. Write the handover in English, regardless of the language of the project. Finally: a short message to the owner – what you found, which fields are open, which secrets/rights issues must be resolved before publishing.