AI working log

How this pack was produced, one row per step. The brief invited AI-native work; this is what that meant in practice. Where a tool made a proposal, the decision stayed with me.

Step	What the AI did	What I decided or checked	Where I did not trust it
Read the brief	Listed the need behind each workshop note and what breaks if only the surface ask is built	Confirmed the list and added the cases the notes never mention: co-borrowers, the second channel, the outage, retention of new data	The brief went in verbatim, never from memory
Decision list	Turned every trap with two defensible answers into Option A / Option B, 14 decisions	Skimmed and kept all 14	The list, not the notes, is what the panel answered
Blind panel	Three independent AI lenses (a delivery analyst, the Bank's compliance and CRM owner, a platform architect) answered all 14, none seeing the others	Read all three	Agreement across blind lenses counted as signal; disagreement went to me
Judge	Merged the three into one proposal sheet: 11 unanimous, 3 split, each split as a branch-day scenario with a recommendation	I made the 3 calls: branch wall with a back-office duplicate check, pending draft on CRM outage, retention stamp at go-live	The judge dropped one lens's self-contradiction and said so
Flow maps	Built current state, target state and failure modes from the decided spec	Read every render as an image	A build that returns no error proves nothing; the picture was checked
Stories	Drafted 12 stories with acceptance criteria in a fixed format, each citing a workshop note and its decisions	Cut and re-cut for length	A cut for length dropped 6 decision clauses; a fresh review caught them, so every cut is now followed by a review
Integration notes	Data flows, the 3 CRM calls step by step, failure modes, questions for the CRM team	Checked every number against the stories	A write-back path had crept in that contradicted a decision; removed
Conformance review	A fresh reviewer, blind to the author, checked every artefact against the decisions, the format and the brief until it passed	Applied the findings	Reviewers could not re-open decisions; anything needing a decision came to me
Hostile review	Three reviewers attacked all acceptance criteria as a squad developer, the DPO and the CRM team lead: 115 findings, 76 distinct issues	64 fixed, 7 defended with a reason, 5 escalated to me	The reviewers found holes; they never rewrote
Escalations	Each escalation as two options with a recommendation	I took all five: flag chip only after selection, a release that survives the next re-read unless the flag changed, an interim monthly deletion report, Bank-named roles for the alert and the merge queue, a subject-access export at go-live	The recommendation was mine to overrule
Two more calls	I questioned the phase-two search fields; the AI ranked identifiers by uniqueness, stability and availability in the room	Exact keys find the person (customer number, IBAN, ID document number); fuzzy fields confirm (name, date of birth, place of birth as an optional narrower); address is never a search input	Every artefact was re-checked after each call
Prioritisation	Applied the phase rule to all 12 stories, collected 28 assumptions with owners, 11 questions for the CRM team, 8 for the Bank	Confirmed the split	Every gap is a labelled assumption with an owner, never a filled-in guess
Sync check	Three independent checks: decisions against artefacts, artefact against artefact, a read as the panel would	Applied the fixes and rebuilt the pack, the deck and the maps	Counts and conditionals drift after late decisions; the check exists for that

Reusable tooling built for this case

Each repeated job became a written recipe an AI tool runs the same way every time. The recipes outlive the case: the next epic starts from them, not from a blank page.

Tool	Purpose	What it fixes in place
Design system recipe	Every artefact looks native to Oper: typeface, colours, card and button rules, copy voice, verification steps, all extracted from the public website and cited to the source	Consistent look across pack, deck and maps without redesigning each one
Flow-map recipe	Turns a decided spec into interactive maps: swimlanes, nodes in a shared grid, hop-by-hop flows, status tags for gap, new, assumed and phase two, a build and a render check	Current state, target state and failure modes as one live deliverable, walked in the session
Epic method	The end-to-end method: brief in verbatim, traps, decisions as options, three blind lenses, a judge, human calls on the splits, artefacts in fixed formats and word budgets, a review loop, a hostile pass, the pack and the deck, this log	Two runs a month apart read like the same analyst; judgment stays with the human at fixed points
Conformance reviewer	A fixed-role reviewer, blind to the author, that checks one artefact against the decisions, the format and the brief and reports block, fix or ok with the location	Catches drift after every edit; it may not re-open a decision
Hostile reviewer	A fixed-role attacker that reads acceptance criteria as a squad developer, a DPO or a CRM owner and reports "this fails when"	Finds holes before a client does; it never rewrites and never decides

What these tools are not: they do not decide. Every split, every escalation and every scope change in this pack was a human call, recorded with its reason.