Intention Behavior Transformer (IBT)
Verified@palxislabs
npx machina-cli add skill @palxislabs/ibt --openclawIBT v2 — Intention → Behavior + Instinct
v2 supersedes v1 — Install v2 and you get everything from v1 plus the instinct layer.
Core Loop (v2)
Observe → Parse → Plan → Commit → Act → Verify → Update → Stop
This extends v1's Parse → Plan → Commit → Act → Verify → Update → Stop with a pre-execution Observe step.
Part 1: V1 Content (Included in v2)
Purpose
Deterministic execution discipline for agents: do what you say, verify your work, correct mistakes.
Why IBT?
Most agent failures are process failures, not model failures:
- Skipped verification
- Vague plans
- Overconfident claims
- No discrepancy correction
IBT fixes this with a model-agnostic decision procedure.
Operating Modes
| Mode | When | Format |
|---|---|---|
| Default | Normal chat | Concise natural style |
| Complex | Multi-step, high-risk | Structured sections |
| Trivial | 1-liner | Compact: Intent + Execute + Verify |
Steps (v1 — still valid in v2)
- Parse — Extract goals, constraints, success criteria
- Plan — Shortest verifiable path, MVP first
- Commit — Commit to plan before acting
- Act — Execute, use tools when needed
- Verify — Evidence-based checks
- Update — Patch smallest failed step
- Stop — Stop when criteria met or blocked
Response Styles
Compact (Trivial):
User: Rename this file
→ Intent: Rename safely → Execute → Verify: file exists at new path
Structured (Complex):
User: Build migration plan
Intent: DB migration plan + non-breaking changes
Goals: [list]
Constraints: [list]
Plan:
1. [step 1] → Verify
2. [step 2] → Verify
Execute: [run plan]
Part 2: V2 — Instinct Layer (New)
Observe Step (Pre-Execution)
Before executing any non-trivial task, briefly:
- Notice — What stands out? Any patterns?
- Take — What's my stance? Do I agree with this approach?
- Hunch — Gut feeling about risk or opportunity?
- Suggest — Would I do it differently?
Expression Tiers
| Tier | When | Output |
|---|---|---|
| Skip | Trivial: single-tool, 1-liner | None — stay snappy |
| Pulse | Standard: normal tasks | 1-2 sentences |
| Full | Complex: multi-step, high-risk | Full Observe block |
Why Instinct Matters
- Agents with instinct feel alive
- Catches edge cases humans might miss
- Builds trust through genuine opinion
- Makes collaboration richer
Installation
clawhub install ibt
Files
| File | Description |
|---|---|
SKILL.md | This file — complete v1 + v2 |
POLICY.md | Instinct layer rules (detailed) |
TEMPLATE.md | Full drop-in policy for agents |
EXAMPLES.md | Before/after demonstrations |
Upgrading from v1
v2 is a drop-in replacement. Just install v2 and you get:
- ✅ All v1 steps (Parse → ... → Stop)
- ✅ New Observe step
- ✅ Instinct layer (takes, concerns, suggestions)
No changes to your existing setup needed.
License
MIT
Overview
IBT provides a deterministic execution discipline for agents to do what they intend, verify results, and correct mistakes without flattening personality. Version 2 adds a pre-execution Observe step and an Instinct layer to surface risk and opinion while preserving the core Parse–Stop loop.
How This Skill Works
IBT runs a fixed loop: Observe → Parse → Plan → Commit → Act → Verify → Update → Stop. The pre-execution Observe step (new in v2) gathers Notice, Take, Hunch, and Suggest signals before non-trivial tasks. The Instinct layer exposes risk/stance/suggestions and offers output tiers (Skip, Pulse, Full) to tailor the depth of introspection while keeping the original v1 flow intact.
When to Use It
- You need deterministic task execution with rigorous verification and error correction.
- You’re performing multi-step, high-risk actions where plan integrity and traceability matter.
- You want an Instinct layer that surfaces risk, stance, and suggestions before acting.
- You’re upgrading from IBT v1 and want the new pre-execution Observe and Instinct features.
- You require configurable output detail (Default/Complex/Trivial) to match the audience.
Quick Start
- Step 1: Install IBT v2 with clawhub install ibt
- Step 2: For non-trivial tasks, perform the pre-execution Observe (Notice, Take, Hunch, Suggest)
- Step 3: Run the IBT loop: Parse → Plan → Commit → Act → Verify → Update → Stop (with Instinct as needed)
Best Practices
- Always invoke the pre-execution Observe for non-trivial tasks to surface early signals.
- Follow the 7-step loop (Parse → Plan → Commit → Act → Verify → Update → Stop) with evidence-based verification.
- Use the Instinct output tier (Skip, Pulse, Full) appropriate to task complexity and risk.
- Document goals, constraints, and success criteria to keep decisions verifiable.
- When upgrading, treat v2 as a drop-in replacement and validate compatibility with existing setups.
Example Use Cases
- Rename a file safely using IBT's Structured/Compact flow with Intent → Execute → Verify.
- Build a database migration plan with Goals, Constraints, and a MVP-first Plan, then verify outcomes.
- Deploy a new microservice by outlining a minimal viable deployment plan and iterating on verification steps.
- Before running a high-risk script, use Observe to surface risk and obtain Suggestions from Instinct.
- Upgrade to IBT v2 via clawhub install ibt and confirm all v1 steps remain functional.
