Menu
Use when experiments are running or just completed, or user shares results and wants to decide what to do next.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Based on SOC occupation classification
| name | execution |
| description | Use when experiments are running or just completed, or user shares results and wants to decide what to do next. |
Read the active session:
ls -t .research-loop/sessions/
cat .research-loop/sessions/<latest>/lab_notebook.md
Summarize the state:
"You're in session [slug]. [N] experiments run. Best so far: [metric]. Last decision: [continue/pivot/kill]."
Create todos with TodoWrite:
Task 1: Annotate run #N — record result, mechanistic explanation, decision
Task 2: Update knowledge graph
Task 3: Write conclusion paragraph (before checking if result matched prediction)
Task 4: Decide — continue / pivot / kill
Read the latest result:
tail -1 .research-loop/sessions/<slug>/autoresearch.jsonl | python3 -m json.tool
Ask the researcher three questions — one at a time:
Q1:
"What happened? Walk me through the result — metric value, direction, was it what you expected?"
Q2:
"Why do you think it happened? I want a mechanistic explanation, not 'the model improved'. What did the change actually do?"
Q3:
"What does this tell you about the next step?"
Append the full exchange to lab_notebook.md:
## Run #N — <node name>
Date: <date>
Mutation: <what changed>
Result: <metric value> (Δ <delta> from baseline)
Researcher explanation: <their answer to Q2>
Causal annotation: <your synthesis of why>
Decision: <continue / pivot / kill>
Next mutation rationale: <why>
Also append a node to knowledge_graph.md:
## [node name] → [result] → [next]
- Mutation: <what changed>
- Result: <metric> Δ<delta>
- Why it worked/failed: <mechanistic>
- Implication: <what to try next>
Apply this tree — ask the researcher first, then give your recommendation:
Improved in last 5 runs?
├── YES → continue this direction
└── NO
├── > 10 runs total with no improvement?
│ └── YES → KILL. Update status, move to next hypothesis.
└── NO → PIVOT. Suggest a different mutation direction.
Show your recommendation explicitly:
"My recommendation: [continue/pivot/kill]. Here's why: [one sentence]."
Update lab_notebook.md status:
## Status
<date>: Run #N complete. Decision: <continue/pivot/kill>. Reason: <why>
Declare success when ALL of these are true:
Then say:
"You have enough to write the paper. Run
/writeor load thewriting-papersskill."
Use when user mentions autonomous iteration, metric-driven optimization, $research-loop plan, $research-loop debug, $research-loop fix, $research-loop security, $research-loop ship, $research-loop scenario, $research-loop predict, $research-loop learn, $research-loop reason, $research-loop probe, or mentions "research-loop" with a goal/metric. Autonomous Goal-directed Iteration — apply Karpathy's autoresearch principles: modify, verify, keep/discard, repeat. Supports bounded mode via Iterations: N inline config.
Mandatory activation layer — loads on any conversation start. Establishes skill-loading protocol, Red Flags, priority rules, and HARD-GATE enforcement for all research-loop skills.
Run a thorough, multi-phase deep research investigation on a topic with subagent dispatch, provenance tracking, and integrity verification.
Experiment sandbox execution for Research Loop. Supports four modes: local (venv), Docker (isolated containers), SSH remote (GPU compute on servers), and Colab (Google Drive bridge). Provides experiment harness templates, code validation, metric collection, deterministic seeding, and compute budget enforcement. Use before running experiments generated by the paper-pipeline.
Use when user wants to explore a topic, find papers, map a field, or understand the research landscape. Not for explaining concepts — use learn for that.
Publication-quality figure generation for research papers. Decision agent selects figure type (code plot vs architecture diagram). Generates Matplotlib/Seaborn code for quantitative figures with iterative improvement loop. Style-matches conference templates (NeurIPS, ICML, ICLR). Use when the paper-pipeline reaches the figure generation phase, or when a user requests figures for an existing draft.