antahkarana
Multi-perspective reasoning through cognitive voices
SKILL.mdskills/antahkarana/SKILL.md
GitHubAntahkarana
Status as of 2026-09-13: trimmed to the voices, the constraints, and the output.
The step list and the per-skill model table were removed — routing is policy in
CLAUDE.md, enforced by hooks/pre-tool-hook.sh.
For the philosophical basis of these six voices — why they're structured this way and how they emerge from retrieval design — see [[../vedanta/antahkarana]] in the vedanta skill graph.
Use when one question needs genuinely different viewpoints: a decision with no obvious right answer, or an approach you're stuck on. Distinct from yajña, which coordinates tasks rather than perspectives on a single question.
The voices
| Voice | Asks |
|---|---|
| manas | What feels right? Quick, practical intuition |
| buddhi | What does the evidence say? Analytical |
| ahamkara | What could go wrong? Risk-aware, protective |
| chitta | What worked before? Memory and pattern |
| vikalpa | What if we tried…? Creative, exploratory |
| sakshi | Neutral witness; synthesizes the rest |
What must hold
- Voices run in parallel and independently. A voice that has read another's answer is no longer a separate perspective, and the whole value is independence.
- Each voice writes to chitta tagged
thread:<id>,voice:<name>, so the synthesis reads from memory rather than from a context window that may have dropped an early voice. - Sakshi's synthesis stays in the orchestrator. It needs every voice at once.
- Report divergence, don't resolve it away. Where the voices disagree is the finding. A synthesis that reads as unanimous has usually just lost information.
Output
Each voice's position in a line or two, then the synthesis: where they converge,
where they genuinely conflict, and a recommendation that says which risk it
accepts. Bracket the run with narrate(action=start|end) so the thread is
resumable.