By name: hdd description: Hypothesis-Driven Development overview and workflow disable-model-invocation: true allowed-tools: Read, Grep, Glob, Bash argument-hint: "[phase]"
Hypothesis-Driven Development (HDD)
Goal: Improved software. Papers are documentation of validated improvements.
HDD is a feedback loop where experiments validate enhancements before merging:
IDEA -> HYPOTHESIS -> EXPERIMENT (on branch) -> RESULTS
|
VALIDATED -> MERGE branch -> improved software -> PAPER
REFUTED -> learn -> NEW IDEA -> iterate
The Critical Feedback Loop
Every experiment runs on a feature branch. Based on results:
| Outcome | Action | Branch | What We Learn |
|---|---|---|---|
| VALIDATED | Merge to main | git merge |
Enhancement works |
| REFUTED | Don't merge | Archive | Why it failed |
| NEEDS-MORE-DATA | Keep branch | Continue | Need larger N |
The HDD Cycle
Phase 0: Discovery
- Capture Idea:
/idea "Research question" - Literature Review:
/literature IDEA-R-XXX "topic" - Formalize Hypothesis:
/hypothesis PAPER-XXX "claim" --target ">=85%"
Phase 1-4: Build-Measure-Learn
- Create Branch + Implement:
git checkout -b exp-XXX-description - Design Experiment:
/experiment H{paper}.{n} "design" - Run Experiment: Collect data
- Decide: Merge (validated) or Learn (refuted)
- Document: Update files, optionally write paper
yurtle-kanban HDD Board
# Add HDD board
yurtle-kanban board-add research --preset hdd --path research/
# Create items
yurtle-kanban create --board research idea "Research question"
yurtle-kanban create --board research hypothesis "Testable claim"
# View board
yurtle-kanban board research
Key Principles
- Software First: Working code, not papers, is the goal
- Branch Per Experiment: Isolate until proven
- Merge Only Validated: No unproven changes
- Learn from Failures: Refuted hypotheses inform new ideas
Related Skills
| Skill | Purpose |
|---|---|
/idea |
Capture research idea |
/literature |
LLM-assisted review |
/hypothesis |
Formalize testable claim |
/experiment |
Design and run experiment |