Adaptive decision framework for AI agents. User checkpoints at every phase.
Works with any AI coding assistant that supports Agent Skills
AI agents jump to implementation. Most failures come from building the wrong thing. AWRSHIFT makes your agent research, define metrics, factcheck, and test in a sandbox — all before touching your main project. One dynamic flow adapts to any task complexity.
| v1.0 | v2.0 |
|---|---|
| 3 modes (Quick/Standard/Scientific) | 1 dynamic flow — scope adapts per phase |
| Text-based questions | AskUserQuestion — structured A/B/C/D choices at every checkpoint |
| No metrics phase | EVALUATE-DESIGN — mandatory success criteria before planning |
| Factcheck only in Scientific | FACTCHECK — mandatory for all scopes |
| No sandbox rules | 10 safety rules — experiment sandbox is isolated from main project |
| No implementation gate | Double gate — DECIDE(GO) + file-by-file preview before touching main project |
One flow. User controls depth.
IDENTIFY → RESEARCH → EVALUATE-DESIGN → HYPOTHESIZE → PLAN → FACTCHECK → TEST → DECIDE → [IMPLEMENT]
At every phase transition, the agent:
- Tells you what was done
- Explains what happens next
- Asks you to choose (A/B/C/D or your own direction)
You're always in control. The agent never proceeds silently.
Claude Code:
mkdir -p .claude/skills/awrshift
curl -sL https://raw.githubusercontent.com/awrshift/skill-awrshift/main/SKILL.md \
-o .claude/skills/awrshift/SKILL.md| You say | AWRSHIFT does |
|---|---|
| "Let's think this through" | Starts IDENTIFY — asks structured questions one by one |
| "Research first" | RESEARCH phase — generates questions, asks you to validate, dispatches agents |
| "Compare approaches" | HYPOTHESIZE — names options, presents comparison table |
| "What metrics should we use?" | EVALUATE-DESIGN — proposes measurable success criteria |
| "Factcheck this plan" | FACTCHECK — verifies plan against original context + optional Gemini |
| "Experiment on [topic]" | Creates experiment folder, starts full flow |
Every experiment creates persistent documentation in your project:
experiments/{NNN}-{short-name}/
├── PLAN.md ← Status, phases, metrics, decisions
├── research/
│ └── 01-{topic}.md ← Agent findings
├── factcheck.md ← Verification results
└── [artifacts] ← Code, configs, outputs
During experiments, the agent NEVER modifies your main project files:
- All work happens in
experiments/folder - Main project files are read-only (for context)
- Only after DECIDE(GO) + your explicit approval → changes proposed to main project
- You see exact file list before any modification
| Skill | When | Purpose |
|---|---|---|
| brainstorm | HYPOTHESIZE phase | Multi-model ideation (Claude x Gemini) |
| gemini | FACTCHECK phase | Cross-model verification |
Both optional. Framework works standalone.
- User-in-the-loop — AskUserQuestion at every phase transition
- Metrics before planning — define success criteria before building
- Factcheck before testing — verify plan against evidence
- Sandbox first — test in experiments/, implement later
- Evidence-based decisions — GO/NO-GO with measured metrics
| Platform | Install |
|---|---|
| Claude Code | Copy SKILL.md to .claude/skills/awrshift/ |
| Codex CLI | Copy SKILL.md to .openai/skills/awrshift/ |
| Gemini CLI | Copy SKILL.md to .gemini/skills/awrshift/ |
| Cursor | Copy SKILL.md to .cursor/skills/awrshift/ |
- claude-starter-kit — ready-to-use project structure with memory, hooks, skills
- skill-brainstorm — 3-round Claude x Gemini adversarial dialogue
- skill-gemini — Gemini toolkit for second opinions, images, diagrams
MIT — see LICENSE for details.
Think before you build. Research before you code. Decide with evidence.