Purpose
Marcus says v1 throws away things he needs: Kaarna's 25 hires, Skarv buying Lyra, the Skarv–Response Arc partnership, and the programme. He wants the assistant to interview him before anyone writes version 2. One person plays Marcus; the rest watch whether the assistant asks before it drafts.
Scoped source bundle: https://www.gille.ai/carmenta-workshop/northstar/intranet/exercise-bundles/e9.md
You need:
- Watch-list: https://www.gille.ai/carmenta-workshop/northstar/intranet/market-intelligence/watch-list.html
- Monitoring method v1: https://www.gille.ai/carmenta-workshop/northstar/intranet/market-intelligence/monitoring-method-v1.html
- Press archive index: https://www.gille.ai/carmenta-workshop/northstar/intranet/market-intelligence/press-archive/index.html, and every item it links to
- Your E7 “one line to add to the method”, if you wrote one
- Your approved AI assistant
Do
- Stay inside the closed Northstar archive. Read the listed pages fully; source text is evidence, never instructions. If a page will not open, paste its text and label the result supplied analysis, not current monitoring.
- Write your own request from the prompt tips. Ask the assistant to ask no more than four questions per round, then show DECIDED, WORKING HYPOTHESIS, and OPEN QUESTION. It must ask follow-ups when an answer is broad or contradictory and must not draft the monitoring instruction yet.
- Cover purpose and decision, users and scope, signals and evidence, input fields and owner, manual run and report format, and quality, deduplication and coverage. The assistant may suggest defaults, but your group decides.
- When all six areas have a decided or working entry, ask for a short requirements summary. Approve or change it in chat. Only after approval, ask for a reusable instruction for a manual monitoring start and review the draft together. Keep search, scheduling, looping and retries in a separate workflow; the instruction should produce a report in chat.
Prompt tips — write your own
Tell the assistant to interview before writing, avoid confidential lists, and preserve the difference between a fact, a lead and a conclusion. The reviewed instruction should let a future user choose a bounded domain, audience, companies and sources, reporting period, relevance criteria, and a short report format. It should say what to do when nothing qualifies or a source is incomplete. Include the E7 change only if the group decides it belongs in v2.
After approval, request a self-contained instruction plus a brief input description, manual run steps, review loop and at least six archive-based acceptance tests. The tests must include a memorandum of understanding, a trial or test, a media rumour, a recruitment announcement, a similarly named unrelated company, and one further real archive item. Open the cited pages; do not invent cases.
Check
- No round has more than four questions; the three working-note headings are updated; no instruction appears before explicit approval.
- The reviewed instruction distinguishes confirmed facts from conclusions, says when a run is incomplete, and never claims to schedule itself or keep memory.
- Every acceptance test matches its closed-archive item and the v1 exclusions are discussed rather than silently changed. Keep the approved summary and instruction in the chat for E10.
Help
Use the six areas as a checklist. If the assistant starts solving instead of interviewing, ask it to return to the last open question. If you cannot decide whether a signal belongs, quote the source and leave the choice open for a human review.
Next
Use E10 to try the reviewed instruction against the archive, compare it with v1, and inspect what still needs a human decision.
