Two weeks almost entirely about identity — and then, underneath it, a quieter question about whether your own records can be trusted.
try? callsSessions per day
Weekends carried 26% of the fortnight — the sturdiest signal here. Wednesday's 34% is one day, not a habit: the other Wednesday was empty. Two samples per weekday is too thin to call a rhythm; it gets real once a few dispatches stack up.
View as table
| Date | Day | Sessions |
|---|
What verification found
Half of what your plans call open is already finished.
View as table
| Verdict | Count |
|---|---|
| Resolved | 4 |
| Partial | 3 |
| Genuinely open | 1 |
You keep building the hard 90% and stopping just before the cheap 10% that makes it land.
What you built
3 headline · 5 more · 6 fixedEvery AI in the app now advises through your own rules
Sit through the Contract interview once — objective function, framing rules, the things that demotivate you — and it flows into expert meetings, Mira, and fresh-project Foundation generation.
This is the moment the Dossier stopped being something you read and became something the app consults. It is also the most opinionated thing in the product: you built a system that takes instruction from you about how to speak to you.
Portfolio values synthesis
Reads five Foundation layers across every project and separates well-evidenced claims from thin ones — only the evidenced ones travel into exports.
Mockup Lab tells the truth now
Imported files become read-only references you fork. Before this, iterating never touched what Claude Code wrote.
Build a Package
Check off exactly the sections and projects you want instead of choosing between four fixed dumps.
/cr-review
Argue with the work card-by-card instead of forcing design opinions through a pass/fail QA frame.
- 426 plans — every plan from the global directory had its path recorded twice, silently killing checkbox write-back
- Expert prompts — three experts fell through to a generic default on any machine but yours
- Note context — attached files and blueprints were dropped from planning prompts on three surfaces
- The guided demo — was overwriting your real creative brief
What it revealed
6 patternsTwo of these come with a chart because the shape is the argument.
Notes completed, by the month they were captured
View as table
| Cohort | Eventually completed |
|---|---|
| February | 88 |
| March | 18 |
| April | 6 |
| May | 1 |
You built four things that require you to remember them
.onAppear. The Claude Code sync writes on enable and never after. Chronicle search stops at the archive boundary without saying so.Still open
7 confirmed · triagedEvery item here was checked against the code, not read off a checkbox — and then sorted by what to actually do about it.
Fix before anything else
How I verified it
ChronicleManager.swift:563 and :177.Finish the last wire
watchOutFor appears nowhere in the detector — so the alerts can misquote your own stated goals back at you.relatedDecisions exists, is documented as the surfacing hook, and has zero callers repo-wide.Decide, don't build
Already done
nothing to do hereRecorded as open, verified as finished. These need no action beyond a store re-sync.
The Dossier's four biggest “remaining” tasks, and 16 of 17 Mockup Lab tasks
All shipped, wired, and confirmed by call site rather than by commit message. The plans simply never got updated.
- Portfolio Pattern Report agent
- DossierContextBuilder injection ×3
- Values section + provenance
- Claude Code projection auto-write
- Mockup import + HTML round-trip
- Opt-in CLI streaming + bubble
Suggested tasks
24 in 7 batchesGrouped so you can hand off a whole batch at once. Items paired to a Still Open card above are already counted there.
Make your records trustworthy again
Turn your best instinct into a system
/cr-silent — a repeatable silent-failure sweep…5 more batches (Design register · Make the app tell you · Backlogs that never drain · Outside eyes · Blueprints)
Loose threads
8 triagedClose — events already answered these
Promote — smaller than their age suggests
What this read
- Nothing truncated. 38 records, well under the 150 cap; no range clamp.
- 8 tasks verified against code, not checkboxes. ~266 others were not — treat that count as unreliable until the stores re-sync.
- Feature and design refinements are proposals from a forward-looking pass, not verified findings.