Generate distinct ideas. Choose with intent. Develop what survives.
One human-in-the-loop creative plugin for Codex and Claude Code.
English · 한국어 · 日本語 · 简体中文 · Español · Français · Deutsch · Português (Brasil)
Imagination keeps the part AI should not automate: the decision. It first creates a portfolio of genuinely different directions, waits for you to choose, then pressure-tests only the selected idea.
flowchart LR
A[Your brief] --> B[3–5 distinct directions]
B --> C{You choose}
C --> D[Pressure-tested concept]
D --> E[Planning]
curl -fsSL https://raw.githubusercontent.com/djfksjd/imagination/main/install.sh | bashThen ask:
Use $imagination to invent several negotiation mechanics for a narrative game
without dialogue trees, hidden dice, or a persuasion stat.
The first response ends with a choice. Reply with a number or direction:
Develop option 2.
| Skill | Best for | Returns |
|---|---|---|
$imagination |
Most users | Portfolio → your choice → developed concept |
$imagination-engine |
Direct divergence | 3–5 useful, non-obvious directions |
$imagination-brainstorming |
An idea already chosen | A decision-ready concept memo |
Important
The router never chooses and develops an idea in the same turn. Automatic chaining reduced constraint fit in testing, so your choice remains the hard boundary between divergence and development.
Fresh preregistered blind comparisons against strong plain prompts:
| Runtime | Preference | Largest measured gains |
|---|---|---|
| Engine v0.5.3 | 50–0 (100.0%) | useful surprise +0.97, diversity +0.66, fit +0.60 |
| Concept Workshop v0.4.2 | 45–5 (90.0%) | actionability +1.37, fit +0.95, causal clarity +0.71, robustness +0.82 |
These are model-judge results from the tested task distribution, not a claim of universal creativity. See the full protocols and limitations in Imagination Engine and Imagination Brainstorming.
# Claude Code
claude plugin marketplace add djfksjd/imagination
claude plugin install imagination@djfksjd
# Codex
codex plugin marketplace add djfksjd/imagination
codex plugin add imagination@djfksjd- Useful surprise: novelty must strengthen the brief, not replace it.
- Human choice: generation and development are separate decisions.
- Constraint fidelity: renamed or indirect violations still count.
- Honest failure: the workshop may reject a direction that does not survive.
- Small runtime: the creative context contains only what improves the result.
The specialist runtimes are maintained and evaluated in their own repositories.
Before a release, synchronize their SKILL.md files, validate all three skills,
run the repository tests, and forward-test the two-turn boundary.
python3 -m pytest tests/ -q
bash -n install.shMIT licensed.