workdesk-configuration

Process unprocessed transcripts in system/intake/ into atlas/meetings, atlas/decisions, atlas/people updates, and gtd/inbox proposals. Uses Gemini 3.1 Flash Lite as the structured extractor (~$0.002/transcript, 4s latency) and Claude for vault integration (wikilink resolution, matching cross-updates, file ops, verification). Operator-confirmed per transcript. Honors matching, action-item ownership, source-documentation, and the source-processing-pattern intake → process → archive flow.

Pull raw verbatim transcripts from Granola (meetings, notes, phone calls), Google Meet standalone Transcript Docs, AND the verbatim Transcript tab inside "Notes by Gemini" Docs — into system/intake/ via three ETL scripts. Pure script orchestration — no AI synthesis, no tokens spent on extraction. Idempotent (skips already-pulled). Default lookback 7 days; flags for 14d / backfill / dry-run / status / single-source. Use when the operator says "get transcripts", "pull transcripts", "grab the last N days of transcripts", or before /process-transcripts.

Unpack complex work or a dense topic one idea at a time, checking comprehension at each step rather than delivering a wall of text. Use when the operator asks to "unpack", be "caught up", or "walked through" something — and offer it proactively before any recap, plan, or explanation long enough to overwhelm in a single message.

Owner-only release management for WorkDesk OS at BenaliHQ/workdesk-os. Pre-merge gate, release cut via scripts/release.sh, post-release smoke test against a pinned canary vault, and rollback. Use ANY time work is being landed into the workdesk-os repo or a release is being cut — downstream operators receive these changes via /update. Trigger phrases — "add this to workdesk", "get this into workdesk", "merge this to workdesk", "ship this to workdesk", "land this in workdesk", "cut a workdesk release", "release workdesk os", "tag a workdesk release", "smoke the workdesk release", "rollback the workdesk release". Also auto-trigger when working in /Users/khalilbenali/code/workdesk-os and the operator says anything about merging, shipping, releasing, or landing changes. Owner-only — not for downstream operators.

Safely start, resume, inspect, and finish isolated repo work when the operator says "work on <repo>", "start working on", "pick up <task>", "parallel session", "is anyone else working on", or "hand this off / I'm done with this task".

Plan, organize, and pressure-test multi-step projects with the GTD Natural Planning Model. A skill family: `/pobo` (guided plan), `/pobo organize` (restructure a brief/brainstorm into clean phases), `/pobo review` (adversarially pressure-test a plan). Use `/pobo --lite` for fast low-stakes planning.

Generate today's daily plan via the daily-plan signal, then surface session-entry items (unprocessed transcripts, intake, due signals). Run first thing in a session, or when context shifts mid-day. Updates state/signals.json after writing.

Five-phase calm orientation for a freshly bootstrapped vault. Silent doctor self-check, welcome, operator-profile interview, brief zone tour, then graduate by either planning a real project with /pobo or starting today's daily note. One question per turn. Idempotent and resumable. Sub-commands --status, --restart, --update-profile.

Produce a shareable, brand-correct HTML deck through a disciplined recon → confirm → build → git-deploy loop. Never improvises brand, repo, or deploy pattern. Use when the operator asks to "make a deck", "build a presentation", "create slides", "build a pitch deck", "build an explainer deck", or similar.

Pulls the latest WorkDesk OS configuration release (skills, rules, scripts, hooks, schema) and applies it to this vault. Does NOT update the WorkDesk desktop plugin/app — that's installed and updated through a separate channel. Walks you through any conflicts in plain language. Operator data (personal/, atlas/, gtd/, intel/, system/) is never touched.

Send feedback, bug reports, and feature requests to the WorkDesk OS team via GitHub. Walks the operator through five short questions like a product manager would, then files the issue. Operator must have the GitHub CLI (gh) installed and authenticated — first run walks through that setup if needed.

Meta-skill — scaffold a new atlas object type via a JTBD-first interview. Object types are structural patterns (book, vendor, deal, etc.) with consistent frontmatter, folder, format, source rule, detection, and matching. Use when an emergent type pattern surfaces or operator says "I want to track X".

Meta-skill — scaffold a new personal/ practice. Practices are operator-owned recurring habits (journal, reading log, gratitude, morning pages). The agent never writes here — practices are read-only from Claude's perspective.

Meta-skill — scaffold a new behavioral constraint at config/rules/{name}.md. Rules are non-negotiable across all skills (no-fabrication, source-documentation, etc.). Use when a correction repeats in ≥2 skills' learnings within 30 days, or when an absolute constraint needs to apply everywhere.

Meta-skill — scaffold a new intel signal type. Signals are Claude-generated synthesis (briefings, observations) drawing from multiple sources through operator-context. Define purpose, anchors, traversal, schedule, output, and detection.

Meta-skill — scaffold a new system source type. Sources are raw inputs (transcript, session-log, intake, bookmark, screenshot, etc.) with a processing rule that turns them into atlas/intel/gtd notes. Define identity, format, processing, retention.

Meta-skill — scaffold a new tool integration (CLI binary, MCP server, or HTTP API) so Claude can use it. Installs it, writes a tool reference, updates the tool note at config/tools/<slug>.md (flips connected:true on smoke-test pass), and links it from the operator profile.

Verifies WorkDesk OS runtime behavior (hooks, locks, latency, session-end export) that filesystem self-check cannot see. Run on first Claude Code session after bootstrap, and any time something feels off. Onboarding pauses until doctor passes.

Materialize a recurring checklist (gtd/recurring/checklists/) as a sequence of actions in gtd/actions/next/. Each step becomes its own action with parent pointing back at the checklist. Use when running a procedure like a publish workflow or expense close.

When Claude is unavailable, hits a hard block, or repeatedly fails the same task, hand it to OpenAI Codex CLI with full vault context. Vault is plain markdown — model-portable. Use after 2+ failed attempts, on persistent API errors, or when the operator says "ask Codex".

Guided spec authoring for builds and deliverables. Produces a spec another agent (or Claude in a later session) can execute against. Use when /pobo produces a deliverable that needs more granularity, or when an operator wants to hand off a build.

Two modes. Default — log the current Claude Code session to system/session-log/ as a structured summary. With --summarize {raw-file} — convert a hook-exported raw session-log file into the final summary note. Either mode preserves the raw conversation alongside the summary.

Upgrade an action in gtd/actions/next/ into a full project folder under gtd/projects/. Re-runs the relevant POBO phases minimally and rewrites references.

Mandatory weekly signal. Generates the weekly-review briefing with proposed closures, promotions, and cleanup. Active from week 1 — the bridge skill between onboarding and steady state. Run at start or end of week, or whenever it's been more than 6 days since the last one.