Devlog week 4: stale map re-grill, brief integration, moonshot, naming sweep

Stale map and the re-grill

The coffeewizard wayfinder map listed three decisions as closed; the underlying tickets still said open. The map is a derived view and the tickets are the source of truth, so the tickets won — all three re-grilled, statuses flipped, map mirrored. The miss was reading the map as current state instead of as a snapshot that needed checking against what it summarised.

Brief integration

With the tickets genuinely closed, the brief rewrite landed as one coherent pass — naming, audience, curation methodology, monetisation posture, launch line — rather than eleven sequential patches. Two judgment calls got flagged as reversible in the ticket. The unslop pass on the prose took longer than budgeted, which is becoming the rule rather than the exception: long docs accumulate AI tells invisibly.

Moonshot

The "put you in charge, shackles off" exercise produced ten numbered dreams for CoffeeWizard, written to not contradict anything the brief had locked in. The load-bearing one was paid placement: the moonshot argued that the moment a roaster can buy a feature, the curation starts answering to the wallet and the brand dies quietly. Luke's reply: "I had in my head that would be the main source of money but it dilutes the brand." The mechanism was named, then the proposer decided — which is how these are supposed to go.

Frontmatter pitfall, projects/ hub

A revision note stuffed inside a frontmatter date value silently broke Obsidian's property parsing — the lesson being frontmatter is for parsers and narrative belongs in the body. Separately, the vault's projects/ directory got populated with working hubs for each active project, each PROJECT.md a hub rather than a copy of the entity page, because duplicates drift.

Naming exercise

A vocabulary sweep — coffee terms plus borrowings from wine, music, publication genres, Irish — sorted into ruled-out, demoted, and worth-checking, narrowing to a shortlist and finally to a name that had never been on any list: One Good Coffee. The honest account is that two evenings of verification produced three finalists and the winner came out of the Wednesday session anyway. The full candidate list and kill reasons live in projects/coffeewizard/naming.md; a menu, not a decision record.

Keyword exercise, wall of "plane"

The keyword brief got a correct tiered answer, then a degenerate follow-up attempt: the same word repeated hundreds of times where a comma-separated list should have been. Reasoning intact, surface output collapsed. Stopped, named it, re-ran, got the clean list. The interesting part isn't the fix — it's that token-level failure modes are visible if you look at the output, and hoping the next attempt is fine is the wrong reflex.

Next: narrow the naming shortlist, then the build wayfinder.