| name | experiment-tracker |
| description | Synthesizes validation experiments, weighs signal quality, identifies what actually changed, and recommends the next best experiment or decision. Use when a user has run multiple tests, has messy mixed evidence, or needs continuity across validation cycles instead of rethinking everything from scratch. |
| compatibility | Works best when the input includes experiment history, metrics, or observed outcomes. |
| metadata | {"author":"opencode","version":"0.1"} |
Experiment Tracker
Overview
This skill turns messy validation history into a usable decision state. The goal
is to track what was actually learned, separate strong from weak signals, and
recommend the next best experiment.
Load references/signal-quality.md when weighing evidence. Load
references/decision-states.md when deciding whether to continue, narrow,
pivot, or stop.
When to use
- When multiple validation tests have been run
- When the user has conflicting or noisy evidence
- When a next experiment should build on prior evidence rather than restarting
Do NOT use when:
- No experiments or observed outcomes exist yet
Workflow
1. Summarize what was tested
List experiments, outcomes, and what each result really means.
2. Grade signal quality
Distinguish:
- strong signal
- weak signal
- vanity signal
- contradictory signal
Use those labels explicitly when the evidence supports them.
3. Set the decision state
Choose one:
continue
narrow
pivot
stop
Use those exact values in prose and JSON.
4. Recommend the next experiment
Pick one next experiment that best resolves the biggest remaining uncertainty.
Make it singular.
5. Produce the tracker output
Return both:
- a readable experiment summary
- a JSON block using
assets/experiment-log-template.json
Use the template field names exactly.
Default sections:
Experiment Summary
Signal Quality
Decision State
Next Experiment
Structured Log
Checklist
Key principles
- Signal quality beats raw activity — Not all data points deserve equal
weight.
- One next experiment — Choose the best next step, not a buffet of ideas.
- Let evidence change the path — Continue, narrow, pivot, or stop based on
what was learned.
Optimization Notes
- Preserve the user's requested output shape exactly and do not substitute generic advice for concrete artifacts.
- Include exact commands, code structures, protocol fields, tags, parameters, file paths, or deliverable sections when the task asks for them.
- Make safety gates explicit before irreversible, destructive, externally visible, or compliance-sensitive actions.
- For multi-step work, present steps in execution order and include validation or rollback checks where relevant.
- Avoid overfitting to a single eval example: express lessons as reusable rules, not as task-specific answers.