Run Exploratory Charters with an AI Note-Taker

Tier: Core · Workflow area: Test design & planning

This playbook wraps session-based test management in AI assistance at three points: charter generation from risk inputs, in-session note capture, and debrief synthesis. The touring, the curiosity, and the bug instinct remain irreplaceably yours — the craft is amplified, not delegated.

When to use

• New or reworked features where scripted cases can't yet exist and human exploration finds what generation can't imagine.
• Post-triage follow-upa risk ranking (pb-06) scored an area 4+ and it deserves human eyes, not just the automated pass.
• Bug-hunt days and release hardeningyou want many focused sessions with comparable, reportable output.

When NOT to use:

• Fully specified verification workif pass/fail is already defined, write the scripted case; chartering it is overhead.
• Sessions where narrating breaks the flowsome deep-focus hunts are better recorded and summarized afterward than annotated live.

Prerequisites

• Risk inputs on handa pb-06 ranking, recent escapes, or at minimum the release diff; charters generated from nothing are generic.
• A capture channel that doesn't steal focusside editor, voice-to-text, or an open chat window; decide before the session starts.
• A session-sheet conventiontimebox, charter, notes, debrief — so sessions are comparable across testers.
• An approved assistant for product content, per your data policy.

The recipe

1. 1.

   Generate charters from real inputs, not vibes. Feed the risk ranking, recent escapes, and the release diff:

   Copy

   Generate 6 exploratory session charters from this context:
   - Risk ranking: [paste pb-06 output]
   - Last 3 production escapes: [paste summaries]
   - Changed in this release: [paste diff summary]
   
   Format each charter: "Explore <target area> with <resources/personas/data>
   to discover <information sought>". 60-90 minute scope each.
   Vary the tour type across charters: data-entry abuse, interruption &
   resume, permission boundaries, back-button & history, slow-network.
   No charter may duplicate coverage the automated regression already has:
   [paste suite area list].
   

   A good charter is specific enough to start immediately:

   Copy

   Explore the refund flow with a partially-refunded order and an expiring
   session, to discover state-consistency failures when authorization
   lapses mid-operation.
   

2. 2.

   Prune and claim. Cut charters that duplicate each other or chase low-value corners; assign owners and timeboxes. You should reject or rewrite at least one — a charter set you accept wholesale is a charter set you didn't read.

3. 3.

   Capture the session as a raw stream. Work the charter; dump observations into a running note as they happen — voice-to-text, a side editor, or messages to the assistant. Don't structure while exploring; structure costs attention the product deserves:

   Copy

   14:12 refund partial 30 of 100, then session timeout forced — order shows
   "Refunded" badge but ledger shows 70 outstanding?? screenshot 3
   14:19 back button after timeout → stale form resubmits, no idempotency key?
   14:26 same flow with admin role — fine. member role only.
   

4. 4.

   Synthesize the debrief with the note-taker. End of session, one prompt over the raw stream:

   Copy

   Turn these raw session notes into a debrief:
   1. CHARTER — restate; note where the session deviated and why that was
      or wasn't justified.
   2. BUG CANDIDATES — each with repro steps as bullets, severity guess,
      and what evidence exists (screenshot #, timestamp).
   3. QUESTIONS — behaviors observed that need a product-owner ruling.
   4. COVERAGE — what was toured, what the charter intended but didn't reach.
   5. FOLLOW-UP CHARTERS — 2-3, derived from what this session opened up.
   Do not upgrade my observations into confirmed bugs. Keep my uncertainty
   markers ("??") visible.
   

5. 5.

   Verify before filing. Re-run every bug candidate's repro yourself before it becomes a ticket — the note-taker faithfully preserves your in-flight misreadings too. Then file through your defect flow (pb-22 turns the evidence into filed-quality reports) and bank the follow-up charters for the next session.

Make it yours

• Voice-first capturedictate observations through your OS speech-to-text into the note; the debrief prompt handles the transcription mess remarkably well, and your eyes never leave the product.
• Paired sessions with a live assistantkeep a chat open during the session and ask for on-the-spot ideation ("what else breaks when sessions expire mid-write?"); log its suggestions as tour extensions but hold the wheel yourself.
• Team bug-hunt scale-upgenerate 15 charters across 5 testers, then feed all five debriefs into one synthesis prompt for a cross-session report: shared themes, duplicate candidates merged, and a heat map of what nobody toured.

Quality check

• Charters trace to inputseach one cites which risk item, escape, or diff area spawned it; no generic "explore settings" filler survived pruning.
• The debrief preserved uncertaintyyour "??" items are still questions, not laundered into findings; any debrief that sounds more confident than your session was is misreporting it.
• Every filed bug was re-reproduced by a human after the session, from the debrief's steps alone — if the steps don't reproduce it, the notes were insufficient and the debrief needs the gap recorded.
• Coverage section is honestunreached charter territory is listed, not silently dropped; it feeds the follow-up charters.
• Session economics holdcapture plus debrief cost under 15 minutes on top of the session; if the AI workflow costs more than it saves, simplify the capture format.