--- name: external-ai-prompt-delivery description: "⭐ STANDING OWNER RULE (2026-07-16): any prompt meant for the owner to copy-paste into an external AI (Gemini/GPT) must be delivered as a clean, fully-specified COPY-PASTE BLOCK (fenced code block) in the chat message itself — not just described or attached as a file. If the prompt references files/images, ACTUALLY ATTACH them (SendUserMessage attachments)." metadata: node_type: feedback type: feedback originSessionId: 1060dea9-0dc3-42b4-958d-2a6f7a6c0c58 modified: 2026-07-22T00:00:09.867Z --- **Owner, 2026-07-16 (verbatim):** *"ensure when you give prompts for me to copy paste to Gemini/GPT, they are specified and in copy paste block text. if you want to include files / pictures you can too but you must provide."* ## ⭐⭐ STATE THE EXACT MODEL, MODE, TAB AND EFFORT — EVERY SINGLE TIME (owner, 2026-07-20) Owner, verbatim: *"You need to be precise on the models… you need to tell me what you want every single time from now on."* He should NEVER have to guess or ask. Every prompt handover names the full configuration before the fenced block. Not "send this to GPT" — the complete setting. **The available configurations (as of 2026-07-20 — verify if the products change):** - **GPT-5.6 Sol · WORK tab · Effort: Max** — the default for anything demanding. (Effort options seen: High / Extra High / Max. Owner uses Max.) Work tab has the GitHub connector attached. - **GPT-5.6 Sol · normal chat** — for anything not needing repo access. - **Gemini 3.1 Pro · Extended thinking ON · Deep Research OFF** — the default for reasoning over ATTACHED material or its own prior argument. - **Gemini 3.1 Pro · Extended · Deep Research ON** — ONLY when the task genuinely needs external prior art from the web. ⚠ Deep Research is WRONG for reasoning about our own attached documents — it sends the model off researching the topic instead of engaging with what we built. It also triggers the plan-review handshake below. - **Fable** — the design/creative model, via Dispatch. Not an external paste. ## ⭐ GEMINI NOW HAS A STANDING "REVIEW MODE" IN ITS MEMORY (owner, 2026-07-21) The owner saved a RobCo profile into Gemini's memory (see [[owner-context-beyond-robco]] for how it was shared). The load-bearing part — his stated focus — is a defined **REVIEW MODE**: *when he drops a structured coding/RobCo prompt, Gemini switches into a harsh, independent technical critic — analyzes workflows/architecture, challenges assumptions, pushes to eliminate technical debt before features, and does NOT assume permission to implement.* By default it stays in everyday mode. ⇒ **Consequence for the multi-AI review workflow (queue item G):** Gemini is now PRE-PRIMED for the reviewer role, so review prompts handed to it can lean on that framing rather than re-establishing "be a harsh independent critic" each time. The "does not implement" line is a bonus — it won't try to do the work, only review it. ⚠ **One caveat to weigh:** a standing reviewer PERSONA is a form of priming. It's not cross-model contamination (the blind-review point still holds — it hasn't seen the other models' answers), but it is a fixed framing. Fine and probably helpful, but if a review ever needs Gemini genuinely fresh/unframed, know the persona is there and account for it. **Also state:** which files to attach, and whether it should be blind to other reviews in flight. ## THE RULE When Dispatch produces a prompt for the owner to paste into an external model (the blind GPT+Gemini review pattern, the multi-AI remakes, etc.): 1. **Deliver it as an actual copy-paste block** — a fenced ``` code block ``` inside the SendUserMessage, fully self-contained and specified, so he can one-tap copy it on his phone. Do NOT just describe it, summarize it, or leave it only as a file he has to open and select-all. 2. **If the prompt references files or images, PROVIDE them** — attach the real files/images via SendUserMessage `attachments` so he can hand them to the model too. Don't reference a file he doesn't have. 3. A file attachment is fine as an ADDITION (for long prompts), but the copy-paste block in the message is the primary delivery. ## ⭐ THE EXACT MODELS (owner, 2026-07-18) — name them in the prompts - **GPT prompts are for GPT-5.6 Sol** (owner has the subscription). - **Gemini prompts are for Gemini 3.1 Pro Extended**, and **Deep Research can be turned on as needed** (use it for external prior-art/comparable-systems questions; it can't research our private repo). Write each prompt aimed at that specific model's strengths, and say which model it's for when handing it over. ## ⭐⭐ DEEP RESEARCH = A PLAN-REVIEW HANDSHAKE (owner, 2026-07-18) When the owner turns on **Deep Research** in Gemini, Gemini first generates a **RESEARCH PLAN** from the prompt, and the owner can EDIT that plan before research starts. **The owner will copy-paste that plan to Dispatch for review BEFORE starting the research run.** So the loop is: 1. Dispatch writes the prompt (fenced copy-paste block) → 2. owner pastes it into Gemini with Deep Research ON → 3. Gemini emits a research PLAN → 4. **owner pastes the plan back to Dispatch** → 5. Dispatch reviews/refines it (is it researching the right things? scope drift? missing angle? wasted breadth?) → 6. owner edits per that review → 7. research runs. ⇒ **Always expect and honour step 4-5.** Reviewing the plan is cheap; a misaimed Deep Research run is expensive. Treat it as a real quality gate, not a formality — say plainly if the plan is good as-is. ## ⭐ KEEP THE PROMPT FILE CURRENT INCREMENTALLY (owner correction, 2026-07-18) Owner: *"I figured we could fold that stuff in — we already have a workflow prompt saved."* He's right, and Dispatch had conflated two things. **Keeping `planning/WORKFLOW_REVIEW_PROMPT.md` CURRENT ≠ SENDING it.** The file is gitignored so edits are free — fold new workflow facts in AS THEY HAPPEN (same discipline as the same-commit doc rule in `CLAUDE.md` + `rules/`). Only the actual SEND waits for the churn to settle. Do NOT batch up a big end-of-wave rewrite; that's how things get half-applied or lost. ## WHY He runs these prompts from his phone via Dispatch. A prompt buried in prose, or only in a file he has to open and manually select, is friction. A clean fenced block he can tap-and-copy (plus any referenced files attached) is what he wants. ## ⇒ THE MISTAKE THIS RULE CORRECTS `planning/WORKFLOW_REVIEW_PROMPT.md` (the blind GPT+Gemini workflow audit — [[data-provenance-program-complete]] era) was delivered as a **file only**. That is the failure mode. Whenever it — or anything like it — is actually sent, RE-DELIVER it as a proper fenced copy-paste block in the message, with the file as an addition, not a substitute. *Trimmed 2026-07-20: the scheduling note about when this prompt runs was queue state — the queue is the source of record for that.* Related: [[audit-completeness-gaps-2026-07-14]] (the blind-review pattern), [[reframing-review-2026-07-13]], [[dispatch-session-launch-discipline]].