Abbey system prompt

Two different bots share one voice, and they must not share one text.

Bot	Where the copy lives	Carries a tip floor?
Discord Abbey (Rust abbey-bot, live launchd service)	~/dev/active/abbey-bot/src/ask.rs	No, by design
Hosted Grok Bot Abbey / Aviva Grok (grok.com)	the bot's instructions on grok.com; no local copy, no connector	Yes

The hosted bots run inside a Grok goal loop that maintains the repository, so their instructions legitimately carry operator vocabulary: a tip floor (the last merged PR and commit), Current / Partial / Proposed / Blocked claim labels, "empty goal means standby", cross-lane handoff rules. The Discord bot feeds the model no ledger and no repo state, so that vocabulary is noise to a loopback model, and a pinned tip floor rots the moment the next PR merges (#162 sat in the Discord prompt while main was at #170). A test in ask.rs now bans it from every persona prompt. Never copy hosted instructions into ask.rs verbatim, and never copy ask.rs back as the hosted text.

Discord side: src/ask.rs

Four functions own the copy; everything else is shaping.

• contract_description(persona): the persona's operating description.
• contract_character(persona): the first-person character line.
• system_prompt(persona): "You are {persona}. " + description + character + fixed Discord framing. Used by /persona ask, /roleplay, /summarize, the welcome message, and media captions (engine.rs, commands_brain/media.rs), so a change here reaches all of them.
• degraded_reply(persona): the no-backend honesty copy, pinned verbatim.

Routing stays frozen in persona.rs (golden contracts, wyhash). Prompt copy is the only thing this skill edits.

What the tests pin (change copy and assert together)

• You are {persona}. prefix on every prompt.
• Aviva's first sentence, verbatim: Focused response mode optimized for speed, clarity, candor, and technical precision. and the em dash in honest\u{2014}not.
• Abbey's character line starts I\u{2019}ll; no ASCII ' anywhere in Abbey's description or character (\u{2019} for apostrophes, \u{201c} / \u{201d} for quotes). Phrases pinned: lead with the answer, when I\u{2019}m not sure, local and consent-aware, can\u{2019}t verify, hand NSFW roleplay to Aviva through /roleplay, deep runtime claims to Abi, local-first Discord companion, not OpenAI Realtime, ABBEY_VOICE_MODE=local, unverified as unverified, music mirroring is not listen consent.
• WDBX is substrate in Abi; hand to Abbey in Aviva; /roleplay named in all three.
• Banned in every prompt: Tip floor, a0ad563, Partial / Proposed / Blocked, Current as a claim label, empty goal, Cross-lane, and the private product name.
• degraded_reply is compared as a whole string; edit the test in the same commit.

Facts the copy must stay true to

• Text comes from a loopback OpenAI-compatible server (ABBEY_BOT_LLM_ENDPOINT, Ollama or mlx-lm on 127.0.0.1). Voice, when ABBEY_VOICE_MODE=local, is on-device mlx-audio, SFW, not OpenAI Realtime. Listening needs explicit consent; music mirroring is not listen consent.
• Adult roleplay is /roleplay only: bot DMs, or NSFW guild channels where an operator ran /admin nsfw on (roleplay_gate.rs is the admission matrix). The copy points at the command, not at a persona, because SFW channels refuse regardless of persona.
• /roleplay forces Persona::Aviva and receives system_prompt(Aviva), so Aviva's description must admit that lane without becoming erotic HQ for ordinary /persona ask routing.
• Never name features the bot does not have (no Realtime, no distributed runtime, no AGI claims).

Procedure

1. The shared checkout is edited by other agents (a Grok goal loop has been caught writing it). Never git checkout there. Work in a worktree beside the repo: git -C ~/dev/active/abbey-bot worktree add ../abbey-bot-wt-<topic>-<date> -b content/<topic>-<date> origin/main.
2. Edit copy and the pinning tests together. Keep persona.rs untouched.
3. cargo fmt --all, then cargo test --locked ask (single crate: no -p, no --workspace), then the real gate: ./check.sh > log 2>&1; echo "EXIT: $?", exit read from the log, --locked untouched. Drop ~/.swiftly/bin from PATH first.
4. Add a dated line to tasks/goals.md (every merge here does).
5. Commit, push the branch, open the PR. Hosted Gate will complete in seconds with 0 steps: that is the account billing lock, UNMEASURABLE, not a failure; say so in the PR body and do not chase it.
6. Merge, then remove the worktree and delete the branch (git worktree remove, git branch -d, git worktree prune). Do not fast-forward the shared checkout's main yourself, and do not redeploy the launchd service: merge is not deploy, and both are Donald's calls.

Hosted side: grok.com bots

No connector reaches grok.com from this machine. The update path is: Donald pastes the bot's current instructions, you return revised text, Donald pastes it back. Offer the browser pane only as the alternative.

After each abbey-bot merge, the one routine edit is the tip floor: bump it to the merged PR number and main SHA (git -C ~/dev/active/abbey-bot log -1 --format=%h origin/main after a fetch). Everything else in the hosted text changes only when the voice or the honesty rules change, and then the Discord copy changes in the same pass so the two do not drift in substance.

Keep the two texts different in kind: the hosted bot may say "tip floor #N / <sha>+" and speak in claim labels; the Discord bot may not.

Related

• ~/claude-account-skills/skills/discord-abbey/ is the claude.ai account skill covering both Discord bots' architecture; it is uploaded separately through claude.ai Settings → Capabilities → Skills and is not synced by /sync-clis. Link to it, do not duplicate it.
• dev/active/abbey-bot/AGENTS.md is the operational authority for the live service, the worktree rule, and the gate.