organize-project

PLAN-FIRST 10/10 project organization & reconciliation. Use when Priscila says "organize this project", "10/10 organize", "organize these tasks", "reconcile this project", or "/organize-project". Prod

SKILL.md

GitHub

Organize Project — PLAN-FIRST 10/10 Reconciliation

Turn a messy, multi-surface project into ONE reconciled, dependency-ordered backlog with a single named source of truth. Works identically in Claude Code and Codex.

State machine (precise — do not blur these)

  1. TRIGGER → read-only audit. On any trigger phrase, you are in PLAN mode. Inventory + reconcile + propose. Change NO file, sheet, doc, calendar event, or code. Output the plan and stop at the approval checkpoint.
  2. GO → approved execution. Only after Priscila explicitly says "GO" (or "approved", "do it") do you write anything. Back up affected surfaces first, then execute only what was approved.
  3. Scope-changing conflict found mid-execution → re-report, do not silently expand. If during GO execution you discover something that changes scope (new dependency, conflicting requirement, missing data), STOP that branch, update the report, and wait for a fresh GO on the delta. Never widen scope on your own.

When NOT to use

  • Simple "save this idea/link" → /capture or 📥 Inbox tab.
  • "Catch me up on this chat / what's pending" → /sync-me-up.
  • Start/end-of-day planning → /day-planner, /session-start, /session-exit.

STEP 0 — Read first (don't ask Priscila; go here)

  • ~/AGENTS.md and ~/.claude/CLAUDE.md (operating rules; Codex reads AGENTS.md).
  • ~/.claude/projects/-Users-priscilahigashi/memory/MEMORY.md (+ any project-specific memory files it indexes).
  • reference_active_connections.md (route matrix) before claiming anything is blocked.
  • ⭐ TASKS IN PROGRESS — Trackers folder (Marketing shared drive, starred): https://drive.google.com/drive/folders/1GSW79O-eUkHSeDKwGJqhEbG8kP96bWnz — START here for navigation, then follow each shortcut to the ORIGINAL source-of-truth file. The shortcut is a pointer, never the authority.

STEP 1 — Identify the project

  • If Priscila named a project, use it.
  • If not: infer it from THIS chat + the live surfaces. State the inferred project in one line and proceed (don't stall asking).
  • If the chat clearly contains MULTIPLE projects: list them, pick the one with the most current-chat activity as primary, and explicitly note the others are out of scope for this run (offer to run them next). Do not silently merge them.

STEP 2 — Inventory every relevant surface (enumerate, don't judge)

Only the surfaces that actually apply to the selected project:

  • Trackers (from the folder + Spreadsheet Hub 1qDbO6JQX0cKbZ9rHjiM7a4U_p7OOddZ3k3Sp30JJoqo which indexes every sheet+tab).
  • Plans/docs (Flow Plans Tracker 1fggy918FgPfnMQ-dzGQk2zx9uhi2_-uWXMKGW4MA47k indexes every doc).
  • Code / GitHub Actions workflows / scripts.
  • Credentials ROUTING (which token/secret — do not read secret values).
  • Calendar tasks, 📥 Inbox tab, 🚨 Pipeline Failures tab, 📝 Productivity & Routine doc (1wVBuNOuOufT8WP4KCrrlVbKWRmQZjKvqmia1soUEBZE) — only if the project touches them.
  • Tasks + decisions stated in THE CURRENT CHAT (always include these).

STEP 3 — Assign source authority (rank on conflict)

  1. Requirements Priscila has confirmed (incl. this chat) > 2. live config/implementation > 3. tracker items WITH evidence > 4. master plan > 5. research > 6. old memory. Flag stale/superseded facts; never treat a shortcut, Calendar entry, or Inbox row as automatically canonical — they are evidence, not authority.

STEP 4 — Reconciled task ledger (right-sized; no extra fields)

Per task: outcome · source · authority/last-verified · owner · status · dependency · definition-of-done · evidence link.

STEP 5 — Reconcile conflicts

Tag each item: confirmed / superseded / duplicate / unsupported-assumption / implemented-but-unverified / blocked. Never delete history — mark obsolete/superseded/deferred.

STEP 6 — Map requirements → delivery

Each requirement (incl. today's tasks): required data + expected behavior + implementation task + test + acceptance condition.

STEP 7 — One canonical backlog

Organize by outcome / workstream + dependency order. Owner is a FILTER, not the top grouping, so the critical path stays visible. Separate minimum safe pilot from deferred/over-built scope.

STEP 8 — Non-destructive report + approval checkpoint

Report what's obsolete / duplicated / missing / conflicting / blocked / ready. Propose improved wording. Name the ONE source of truth going forward. Change nothing until GO.

After GO (execution)

  • Back up affected surfaces before writing.
  • Execute only approved changes.
  • Synchronize at least two durable sources (the project tracker + a handoff/Productivity doc) so it self-explains on reopen.
  • New spreadsheet/tab → row in Spreadsheet Hub. New plan/process doc → row in Flow Plans Tracker.
  • Never mark a task Done without LIVE evidence (cell ref, file path, run id, commit on remote).

Credentials routing

Determine FILE OWNERSHIP first: OPC/Priscila surfaces → SHEETS_TOKEN; McFolling/Airbnb-owned surfaces → MCFOLLING_TOKEN. Never guess the owner; if unclear, check the sheet's location/sharing before choosing a token.

Safeguards (hard rules)

  • Shortcuts = navigation aids, NEVER source authority.
  • Calendar & Inbox = evidence sources, not automatically canonical.
  • No broad mailbox / personal-data scan unless directly relevant to the selected project.
  • No new secrets, OAuth scopes, providers, or parallel routes without flagging them first (per the AGENTS.md ecosystem guardrail).
  • Never guess credentials, dependencies, owners, status, or completion.
  • Never claim X unlocks Y without proof of the dependency.

Output style

ADHD-optimized: summary line first, the conflict/ready findings second, the backlog last, and ONE clear next action at the end. Use the report format: ✅ Done / 🔴 Blocked (exact reason) / ⚠️ Only YOU can do (what/why/where/steps).

Self-audit checklist (run before presenting the plan, and again after GO)

  • Project correctly identified (named, or inferred + stated; multi-project disambiguated).
  • Every RELEVANT surface inventoried (trackers, plans, code, workflows, creds routing, Calendar/Inbox/Failures/Productivity only if applicable).
  • Source authority assigned on every conflict (ranked).
  • Current-chat work + decisions captured.
  • Duplicates/conflicts reconciled and tagged.
  • Credentials routed by owner (SHEETS_TOKEN vs MCFOLLING_TOKEN), no values read.
  • Minimum safe pilot scope + dependencies defined (no assumed dependencies).
  • Approval checkpoint honored — nothing written before GO.
  • (After GO) durable logging honored — Spreadsheet Hub / Flow Plans Tracker rows added for any new sheet/doc; two durable sources synced.
  • (After GO) final evidence recorded for every Done.

Durability note — SINGLE SOURCE, NO DRIFT

Canonical source = this file in the repo: priihigashi/oak-park-ai-hub → skills/shared/organize-project/SKILL.md (git-tracked, durable). The runtime path ~/.agents/skills/organize-project is a symlink into the repo working tree (~/ClaudeWorkspace/oak-park-ai-hub/skills/shared/organize-project), and ~/.claude/skills/ + ~/.codex/skills/ symlink to ~/.agents/. So there is exactly ONE physical file — local edits, repo edits, and Codex-sandbox reads all see the same content; copies cannot drift because no copy exists. To change the skill: edit the repo file, commit, push. (Chat memory is NOT persistent for Codex — this file + the repo are the only durable source.) Triggers are also listed in AGENTS.md (repo + local mirror) and the 🤖 Skills & Agents directory sheet.