Fieldmark

Fieldmark is a Flutter app that turns a PDF floor plan into a live installation checklist — foremen drop numbered pins, installers pick up tasks offline in the field.

Flutter · Supabase · Riverpod · Firebase FCM · 2026-09-12 → 2026-09-19

Context

Installer crews (electric, HVAC, low-voltage) work off paper plans and text messages. A foreman marks up a PDF, hands it out, and then chases status by phone all day. Fieldmark makes the same PDF interactive: pins are tasks, tasks have photos and voice notes, status flows back automatically.

Problem

Existing task apps (Asana, ClickUp, Fieldwire) treat plans as attachments. In practice, installers navigate by where on the plan the work is, not by task title. And they lose signal in mechanical rooms. Todo apps don't survive that. Fieldmark starts from the PDF and works offline first.

Division of Labor

Prompt Architecture — Analyze → Propose → Decide → Implement

CLAUDE.md defines a four-stage gate: I state the problem, Claude analyzes and proposes 2–3 options with trade-offs, I decide, then Claude implements. No code before decision. Any change touching more than 2 files or a UX flow gets a written plan first.

The knowledge layer is DEVLOG.md, structured as Problem/Decision/Result/Lesson blocks. It's not a diary — it's the memory Claude loads next session so we don't re-litigate solved problems. .claude/settings.local.json pre-permits 30+ flutter/git commands so nothing stalls on approval prompts.

Iteration Moment — Apple Sign-In rewrite

• Before: 2026-09-12 f05b752 — initial build used Supabase's web-based signInWithOAuth for Apple. Apple review rejected Build 20 with Guideline 2.1(a): sign-in flow broken.
• After: 2026-09-18 8b18221 feat(auth): switch Sign in with Apple to native ASAuthorization (build 21) — pivoted to native ASAuthorization with nonce + sha256 validation. 5 files changed. Submitted the next morning, 2026-09-19 071b10c.

Lesson: When Apple bounces you on auth UX, Claude can implement the native replacement in one session — but only if DEVLOG already contains the rejection text verbatim. Feelings don't route back into code; error strings do.

Trade-Offs

1. Riverpod over Bloc. Less boilerplate, code generation, faster velocity on a solo project. Cost: fewer middleware plugins.
2. Supabase Postgres RLS over Firebase. Real row-level security in SQL. Cost: no built-in offline sync — handled by connectivity_plus + a local queue.
3. Native ASAuthorization over web OAuth. No redirect dance, Apple review passes. Cost: Apple client secret rotates every 6 months — automated with a Python script and a calendar reminder before the first user signed up.
4. PKCE flow. Required for native Apple Sign-In. Cost: nonce validation complexity — mitigated by codegen.

Results (honest)

• 7 days from first commit to App Store submission
• 7 commits, 5 of them in the final 24h auth sprint
• Build 20 rejected → Build 21 submitted → waiting for review as of 2026-09-19
• 0 shipped users at the time of writing — this is the first submission that passed the auth gate

Reflection

1. Rejection emails are the best specs. Guideline 2.1(a) with the reviewer's screenshot was cleaner than any AC I would have written myself.
2. A DEVLOG that captures why survives context resets. Claude Code hits context limits mid-project; DEVLOG is what lets the next session pick up without me re-explaining.
3. Automating the boring future task pays off before the future arrives. The 6-month Apple secret rotation script was written the same day as the auth pivot — because the pivot's context was the only time I'd remember to.