research-director

How to write an agent definition for this project. Use when the user asks to write, add, create, or redesign an agent — even if they don't say 'agent file' explicitly. Also use when deciding how an agent prompt should be structured, what sections it needs, or how to split responsibilities between agents and orchestrator skills. Trigger on: 'write an agent', 'create an agent', 'design the agent', 'update the agent', 'agent prompt for X'.

Exit rubric for the Experimental Setup Module. Load when the Main Agent needs to determine whether EXPERIMENT_SETUP.md is complete enough to hand off to Phase 1 (Method Design). Returns PASS or NEEDS-REWORK with specific failing criteria listed. Evaluates structural completeness and verification quality only — not whether the chosen baselines are the 'right' ones (that judgment belongs to the Architect).

Main Agent orchestration skill for the Experimental Setup Module — the pre-Phase-1 step that audits which datasets, metrics, baselines, LLMs, and reusable code are actually available for the idea. Load when Loop 2 has passed and EXPERIMENT_SETUP.md has not yet been produced, or when the previous audit flagged thin coverage and the Main Agent needs to drive a supplementary scout search and re-audit. Drives the auditor → exit-check → optional supplement cycle and hands off to Phase 1 (Method Design).

Exit rubric for Loop 1 (ideation refinement). Load when the Main Agent needs to determine whether the current IDEA.md is sufficiently developed to proceed to Loop 2 stress testing. Returns PASS or FAIL with specific failing criteria listed. Evaluates development completeness only — not idea quality, novelty, or validity. Those are Loop 2's job.

Main Agent orchestration skill for Loop 1 (ideation refinement). Load when entering or re-entering the idea refinement loop — whether starting fresh with a new idea, responding to user intervention during ideation, or rolling back from Loop 2 after the Tester returned attack points. Drives the Scout → Refiner → exit-check cycle and controls loop termination.

Exit rubric for Loop 2 (stress testing). Load when the Main Agent needs to determine whether the current idea has survived adversarial scrutiny sufficiently to proceed to the Experimental Setup Module. Evaluates the latest Tester report against four criteria. Returns PASS or FAIL with specific failing criteria listed. Evaluates resilience to attack only — not development completeness (that was Loop 1's job).

Main Agent orchestration skill for Loop 2 (stress testing). Load when Loop 1 has passed and the idea is ready for adversarial evaluation, or when re-entering Loop 2 after a Loop 1 retreat fixed the Tester's attack points. Drives the Tester → exit-check → retreat-or-proceed cycle. Controls loop termination and user notification.

Exit rubric for Phase 1 (Method Design). Load after the Architect ↔ Confirmer loop has converged (Confirmer Verdict = ALL CLEAR) and the Main Agent needs to determine whether the confirmed spec is complete enough to hand off to Phase 2 (Small-Scale Verification). Returns PASS or NEEDS-REWORK with specific failing criteria listed. Evaluates structural completeness and dual-anchor alignment only — semantic judgments (clarity, quality) are the Confirmer's responsibility upstream.

Main Agent orchestration skill for Phase 1 (Method Design). Load when the Experimental Setup Module has passed (EXPERIMENT_SETUP.md exists) and Phase 1 has not yet produced a confirmed method spec, or when re-entering Phase 1 after an upstream revision (IDEA.md or EXPERIMENT_SETUP.md changed). Drives the Architect → Confirmer loop with round cap and convergence detection. The Main Agent only monitors — it does not inject its own judgment between rounds. Hands off to Phase 2 (Small-Scale Verification).

Exit rubric for Phase 2 (Small-Scale Verification). Load after the Analyst has written `phase_2_verification/analysis/analysis_{n}.md` against the latest finalized `result_{n}.md`, and the Main Agent needs to determine whether Phase 2 is done (proceed to Phase 3) or whether the run calls for a rollback. Evaluates structural completeness of the analysis only — the routing decision itself belongs to the Main Agent, which reads the dominant rollback depth this check surfaces. Does NOT re-diagnose results, does NOT auto-invoke Loop 3.

Main Agent orchestration skill for Phase 2 — the code-implementation sub-phase that turns the confirmed method spec into runnable code, one module at a time. Load when Phase 1 has converged (method_design_v{n}.md with ALL CLEAR confirmer report exists) and Phase 2 has not yet produced a full set of fidelity-passed modules; also load on Loop 3a re-entry when a specific module needs re-implementation. Drives the per-module Coder ↔ Fidelity Checker loop with a per-module round cap and an outer module-iteration loop. The Main Agent only monitors — it does not inject its own judgment between rounds. Hands off to the (not-yet-built) run-experiment skill.

Exit rubric for Phase 3 (Main Experiments). Load after every declared matrix cell has a finalized result_{n}.md + analysis_{n}.md, and the final story-coherence-advisor checkpoint has produced a story_check_{k}.md. Returns PASS or NEEDS-REWORK with a routing-advisory table (proceed to Phase 4 vs. rollback target). Evaluates structural completeness of the matrix's analyses and narrative coherence signal — NOT the quality of the results, NOT the routing decision itself (that belongs to the Main Agent), NOT ablation design (that is Phase 4). Does NOT re-diagnose any cell's analysis.

Main Agent orchestration skill for Phase 3 — scaling the Phase-2-validated minimum-viable slice up to the full EXPERIMENT_SETUP.md matrix (all datasets × all LLMs × all baselines + method) that a paper's Main Results table needs. Load when phase2-exit-check PASSed with dominant rollback depth None, and Phase 3 has not yet produced a fully analyzed matrix; also load on re-entry after Loop 4 supplementary experiments. Drives: (i) spec upgrade from method_design_v{n}.md to framework_v{m}.md via method-architect ↔ spec-confirmer, (ii) code upgrade via the per-module Coder ↔ Fidelity loop, (iii) matrix execution via multiple run-experiment calls, (iv) per-cell experiment-analyst invocation, (v) story-coherence-advisor checkpoints. Hands off to phase3-exit-check. Does NOT run ablations — that is Phase 4.

Main Agent orchestration skill for Phase 4 — taking the Phase-3-validated framework_v{m}.md and latest story_check and producing per-variant ablation evidence that sharpens contribution attribution. Load when phase3-exit-check PASSed with dominant rollback depth None and a handoff-compatible story check, and Phase 4 has not yet produced a fully analyzed set of variants; also load on re-entry after Loop 4 supplementary ablation additions (Case B) or after a phase4-exit-check NEEDS-REWORK (Case C). Drives: (i) plan design via ablation-designer, (ii) per-variant spec delta via method-architect ↔ spec-confirmer, (iii) per-variant code upgrade via the per-module Coder ↔ Fidelity loop with carry-forward from Phase 3, (iv) per-variant run-experiment + experiment-analyst + story-coherence-advisor Context C checkpoints. Hands off to phase4-exit-check. Does NOT rerun Phase 3 main-results cells.

Exit rubric for Phase 4 (Ablation). Load after every variant in the ablation plan's execution set has a finalized result_{n}.md + analysis_{n}.md, and the final story-coherence-advisor Context-C checkpoint has produced a story_check_{k}.md. Returns PASS or NEEDS-REWORK with a routing-advisory table (proceed to Phase 5 / 6 vs. rollback target). Evaluates structural completeness of the ablation sweep's analyses and narrative coherence — NOT the quality of the ablations' verdicts, NOT the routing decision itself (that belongs to the Main Agent), NOT the paper framing (that is Phase 5 / 6). Does NOT re-diagnose any variant's analysis.

Exit rubric for Phase 5 (Review). Load after both reviewer- constructive and reviewer-adversarial have produced reports for the current round, and the Main Agent needs to determine whether Phase 5 is done (structurally complete) and what the aggregated signal is (Proceed to Phase 6 vs. Loop 4 supplementary vs. halt). Returns PASS or NEEDS-REWORK with a routing-advisory table (Phase 6 vs. Loop-4-target vs. halt). Evaluates structural completeness of the two reviewer reports and their severity aggregation — NOT the correctness of individual concerns, NOT whether Loop 4 is 'really' needed (that is the Main Agent's call with the user), NOT paper- readiness (that is Phase 6's job). Does NOT re-review the research, does NOT arbitrate disagreements between reviewers, does NOT auto-route.

Main Agent orchestration skill for Phase 5 — taking the Phase-4- validated attribution evidence and latest COHERENT story_check, and producing a paired constructive + adversarial review of the research-as-it-stands. Load when phase4-exit-check PASSed with Aggregate rollback depth None (or High absorbed by Criterion 5) and a COHERENT latest story_check, and Phase 5 has not yet produced both reviewer reports; also load on re-entry as Case B after Loop 4 supplementary experiments landed, and as Case C after phase5-exit-check NEEDS-REWORK. Drives two sub-processes: (i) paired reviewer invocation (Constructive + Adversarial — parallel by default, serial on user request), (ii) per-report structural check + user surfacing before handing to phase5-exit-check. Does NOT re-interpret reviewer reports, does NOT write the paper draft (Phase 6), does NOT design supplementary ablations or experiments (those route to Ablation Designer Case B or Phase 3 Case B downstream).

Exit rubric for Phase 6 (Writing). Load after phase6-writing has produced phase_6_writing/drafts/paper_v{n}.md by assembling per-section drafts, and the Main Agent needs to determine whether Phase 6 is done (structurally complete) and what the routing signal is (Proceed to Phase 5 Case C for paper- draft review vs. revise specific sections vs. escalate to Phase 5 Case C anyway for semantic review). Returns PASS or NEEDS-REWORK with a routing- advisory table. Evaluates structural completeness of the paper draft and per-section files, claim-evidence trace fidelity, and anchor consistency (contribution bullets / spec modules / summary tables) — NOT the prose quality (that is Phase 5 Case C's job), NOT whether the contribution bullets are 'right' (that is Story Advisor's job), NOT whether the draft is submission- ready (that is Phase 5 Case C's Overall Recommendation). Does NOT re-draft any section, does NOT arbitrate between Writer choices and reviewer concerns, does NOT auto-route.

Main Agent orchestration skill for Phase 6 — taking the Phase-5- passed review evidence (PHASE_5_SUMMARY.md + both reviewer reports) and producing a paper draft section by section via the writer opus agent. Load when phase5-exit-check PASSed with 'Proceed to Phase 6' routing (0 Fatal + ≤2 Major with Writer-level or Story-Advisor routing) and Phase 6 has not yet produced a full paper draft; also load on re-entry as Case B after Phase 5 Case C raised Writer-level concerns on a prior draft, or after Loop 4 supplementary evidence landed and cascading Phase 5 / Story Advisor updates have re-entered Phase 6. Drives three sub-processes: (i) section plan declaration (order + paper-type conventions), (ii) per-section writer invocation serial in dependency order, (iii) paper assembly by concatenating the latest section versions into phase_6_writing/drafts/paper_v{n}.md. Hands off to phase6-exit-check. Does NOT review the draft (that is Phase 5 Case C), does NOT edit writer output, does NOT design new experiments.

Main Agent orchestration skill for Phase 2 experiment execution — the step that takes fidelity-passed code and produces a result_{n}.md for the Analyst to interpret. Load when Phase 2 implementation has converged (every module's fidelity report is ALL CLEAR or accepted-with-override) and an experiment has not yet been launched; also load on Loop 3a re-entry after a code fix, or in Finalize mode when a prior Launch left result_{n}.md in RUNNING state and the run is now complete. Backend-agnostic: the skill defines the lifecycle contract (Launch / Finalize, sentinel-based completion detection, result file schema) but does not prescribe how commands are executed — local, SSH, SLURM, container, or any other backend the user has configured. Launches in background (via whatever detach mechanism the user's backend supports) so long runs survive /clear. Does NOT classify experiment failures — that is the Analyst's job downstream.

Persist current project state before ending a session or running /clear. Use when the user says 'save', 'wrap up', 'new session', 'ending here', or is about to run /clear. Writes all current state to PROJECT_STATE.md so the next session can resume without ambiguity. Must be run before any /clear or context reset.

Restore full project context at the beginning of a new session. Use when the user says 'start session', 'restore context', 'new session', 'resuming', or starts a conversation on this research project without context — reads PROJECT_STATE.md, reloads phase-specific artifacts, checks for running experiments, and announces the current state before taking any action. Must be run before any pipeline work begins.

How to write a skill for this project. Use when the user asks to write, add, or create a new skill. Also use when deciding how a skill should be structured, what sections it needs, or how to write its description field. Trigger on: 'write a skill', 'create a skill', 'add a skill', 'new skill', 'make a skill for X', 'how should I structure this skill'.