PrettySeriousResearcher: Skill for Claude Code

Research-paper orchestrator

This is the entry point and state machine for building a research paper. It is deliberately thin: it figures out where we are, opens the session clock, routes to the right phase craft skill, enforces a few hard-stop gates, and reports effort + cost. The actual work lives in the phase and cross-cutting skills it invokes.

Optional run focus: $ARGUMENTS (e.g. a phase name, a topic, or "resume").

Lifecycle (effort weights are planning targets, not quotas)

Total effort budget ≈ 120 agent-active hours, where "hours" means the user's interaction time with the coding agent — framing options, confirming at gates, steering — not human research labor and not the agent's autonomous background runtime (that's a separate wall-clock line). Because the research and writing are automated, the calendar cycle collapses; what this budget tracks is time-in-the-loop. See research-tracking for how the two clocks are measured and kept apart.

After Phase 4, the submission tail (research-submission: execute the submission, rebuttals, revise-and-resubmit, camera-ready, release) runs outside the four-phase effort split — it is calendar-gated by the venue, not effort-gated, and every outward-facing step in it is user-confirmed.

On every engagement (do this first)

1. First run? If docs/tracking/state.json has a null first phase_windows[].start and the repo lacks the scaffold (docs/, README.md, architecture.md, .gitignore, memory.md), bootstrap it — create the layout described in research-repo-hygiene and set the first window's start to now. Then begin Phase 1.
2. Open the session clock. Set/refresh docs/tracking/.session (start, last_heartbeat, phase). The authoritative time/token accounting is derived from transcripts by research-tracking; .session just marks the open session.
3. Read state. current_phase and the gate flags from state.json tell you where we are and what's allowed next. If docs/HANDOFF.md exists, read it before architecture.md: it is the rolling handoff (in-flight runs, decisions only the user can make, next steps, environment pitfalls), rewritten at every milestone and before any context reset — a compaction summary is not a substitute (earned: a mid-battery context reset resumed cleanly from it). Before repeating any "pending user action" from the handoff, re-run its acceptance test (a credential probe, a tier check, a file's existence) — the user may have done it since; a stale pending item asked twice costs trust (earned: a data-vendor top-up listed as pending had been completed days earlier).
4. Recall user patterns. Read root memory.md (untracked) for the user's working habits; apply them unless they contradict a project skill (skills win on methodology; memory wins on the user's explicit stated choices — see research-repo-hygiene).
5. Report status. Run research-tracking and surface the compact effort+cost line, then state the current phase and the immediate next step.