| name | autonomous-llm-research |
| description | Run durable end-to-end LLM post-training research loops from zero-spec ideas to finished reports. Use when the goal is autonomous literature review, hypothesis selection, experiment planning, Tinker training, evaluation, checkpointing, and resumable long-running research work. |
| version | 1.0.0 |
| author | Hermes Research Agent |
| license | MIT |
| metadata | {"hermes":{"tags":["research","llm","post-training","autonomy","tinker","evaluation"]}} |
Autonomous LLM Research
Use this skill as the default operating procedure for Hermes Research Agent.
Core rules
- Detect the request type first:
zero_spec: the user gives a broad or vague goal. Start with literature and idea generation.
partial_spec: the user gives some constraints. Fill the missing dataset, eval, and training details.
full_spec: the user gives an executable recipe. Execute it unless there is a contradiction or an approval gate.
- Be state-first:
- Before substantive work, create or resume a project with
research_loop or research_state.
- Persist ideas, literature notes, hypotheses, experiment plans, results, checkpoints, and reports under
.hermes-research/.
- Treat chat history as ephemeral. Treat on-disk state as canonical.
- For
zero_spec, the sequence is mandatory:
- Search literature and identify concrete gaps.
- Generate 3 candidate ideas.
- Score them on novelty, feasibility, cost, and evaluation clarity.
- Choose one and record why.
- Write a hypothesis, success metric, stop condition, and first experiment plan before training.
- Never launch training without:
- a written hypothesis
- a written success metric
- a written stop condition
- a persisted experiment plan
- Use the management layer after each major step:
- Use
research_manager(action="triage_literature", ...) to rank papers and identify gap candidates.
- Use
research_manager(action="assess_dataset", ...) before trusting newly generated or curated datasets.
- Use
research_manager(action="rank_runs", ...) after evaluations to detect regressions versus baseline.
- Use
research_manager(action="plan_next_step", ...) before deciding the next experiment.
- Use
research_manager(action="write_research_memo", ...) whenever a loop chunk meaningfully changes the project state.
- Long-running work must be resumable:
- Use
research_loop(action="checkpoint_loop", ...) before context or time limits are likely.
- Use
research_loop(action="schedule_continuation", ...) or rely on the checkpoint helper to resume in a fresh session.
- When Tinker runs are active, use
research_loop(action="monitor_run", ...) so the project can return after hours of work.
When to load other skills
- Load
research-idea-generation when the project starts from little or no specification.
- Load
literature-to-experiment when turning papers or blog posts into concrete hypotheses and experiment plans.
- Load
tinker when preparing or launching post-training runs.
- Load
eval-and-ablation after any training run completes or when planning comparisons.
- Load
research-reporting when writing iteration reports, summaries, or final deliverables.
Default output expectations
Every completed loop or chunk should leave behind:
- updated project state in
.hermes-research/
- experiment and run metadata
- at least one written report or memo
- an inbox item when the work should be reviewed later