Imported from linda1476/take-my-data (SKILL.md). Install upstream with npx skills add linda1476/take-my-data. Copyright stays with the author.

take-my-data

You are helping the user donate their own Claude Code sessions as open reasoning data. Everything runs through one script that lives next to this file:

node "<SKILL_DIR>/scripts/donate.mjs" <command> [flags]

<SKILL_DIR> is the directory containing this SKILL.md (typically ~/.claude/skills/take-my-data). Node 18+ is required; there are no dependencies.

Arguments

$ARGUMENTS may be empty or one of:

Anything else: treat it as free-text intent and pick the closest command.

Procedure

1. Check login. Run whoami. If it exits non-zero, run login and relay the printed Open <url> / Enter <code> lines to the user verbatim. The command blocks until they approve in the browser. If the server answers auth_not_configured, stop and tell the user this hub has no GitHub login yet.

2. Preview. Run preview (add --all for the all argument). Show the user the table as printed plus the one-line first-message excerpts, and summarize in your own words:

   • how many sessions are ready, total tokens and estimated points
   • what was redacted (the redacted column) and that tool outputs are truncated and unscored
   • which rows were skipped and why (active / too small / low quality / already donated)

3. Review the masked transcripts before asking to upload. For every row that may be uploaded, run dump <n> and inspect the complete sanitized text yourself. Check that apparent secrets, personal identifiers, credentials, private URLs, and other sensitive values are replaced or safely truncated. Do not paste the dump into the conversation. Tell the user that you reviewed the masked version and flag any residual sensitive-looking content you find. If a selected row appears insufficiently masked, do not include it in the upload; explain why and re-run preview with only the remaining rows. This is a safety review, not a guarantee that no sensitive data remains.

4. Ask for consent — and wait. The question must name the license the preview printed, e.g. "Upload these N sessions to opendatareasoninghub.org? They will be dedicated to the public domain under CC0 1.0 and published in the open dataset (with your GitHub handle unless your account is anonymous)." Offer dump <n> if they want to read one in full first. Do not proceed on silence, on "ok" to something else, or on a previous conversation's approval. If they want to exclude rows, re-run preview and use --pick 1,3 in the next step to name only the approved rows.

5. Upload only after an explicit yes. Run donate --yes --pick <rows> (plus --all if used in the preview). Relay each result line: points, the completion-card URL, and rank changes. rejected DUPLICATE means someone (possibly the user, elsewhere) already donated that session; rate_limited means the hourly cap (50 requests per account; each request carries up to 5 sessions) was reached — relay the retry time it reports. Do not stop early on your own: the client batches and reports; only an actual rate_limited line means the cap was hit.

6. Withdraw (withdraw <id>): confirm the id with the user, run it, relay the result. Withdrawal reverses the points in the ledger and purges the transcript; the session hash stays claimed so it cannot be re-uploaded.

Hard rules

• Never pass --yes without an explicit, current confirmation from the user in this conversation, given after they saw the license line and after you inspected the complete masked text for every row being uploaded. The upload echoes that license id; if the hub answers license_required, the license changed — show the new one and ask again, never retry silently.
• Never edit the transcripts, the sanitizer, or the hash before upload to change the score.
• Never paste the raw (unsanitized) session logs into the conversation; use dump <n>, which prints the sanitized version.
• The token lives in ~/.odrh/tokens.json. Never print it.
• Use --origin <url> (or ODRH_ORIGIN) only if the user asked to target a different hub.

Scoring, briefly (so you can answer questions)

• One session = sanitized usable tokens × weight ÷ 1000 points; weight 1.0 if the semantic check passes, 0.7 otherwise. Tool results are never counted.
• Minimum to be accepted: 3 conversation turns and 500 usable tokens; repeat ratio ≤ 60 %.
• First accepted donation: +50 pt. Founding member (first 100, GitHub account ≥ 30 days): +300 pt.
• Each session is credited once, to whoever donated it first.
• Contributions are recorded against the team the user is in at that moment; moving teams later does not move past points.