Cognitive Signature IDE
Cognitive Signature IDE — a Claude Code plugin that syncs Claude to your cognitive signature and keeps it aligned as you change.
Core trick
Extracts style dimensions from past user prompts in Claude Code JSONL logs and injects them into each prompt via a plugin hook.
Description
Cognitive Signature IDE — a Claude Code plugin that syncs Claude to your cognitive signature and keeps it aligned as you change.
Two functionalities on one pipeline:
1. Voice signature — 7 dimensions of how you direct AI (directive style, compression, reframe patterns, trust signals, idiomatic tells, iteration cadence, texture energy). Extracted from raw user-typed directives across your Claude Code JSONL corpus.
2. Operational signature — 4 dimensions of what you've learned through usage (recurring_decision_templates / recurring_failure_patterns / recurring_tooling_invocations / vocabulary_anchors). Auto-promoted to a permanent signature when n≥2 instances observed.
Four onboarding presets (normie / power / team / enterprise) map to three deploy modes (standalone / team-shared / cloud-governed). Hands-off by default; user surfaced only on conflict.
Ships 3 dual-function governance subagents — brutus (Opus, adversarial), qa (Haiku, schema/compile/dead-code), historian (Sonnet, drift tracking) — across two substrates (in-session Claude Code plugin + API-level Managed Agents). Same architecture, different persistence.
Measurement: 100% auto-scorer accuracy on N=10 blind tests (Claude-as-judge) = +66.7pp over chance, 5-persona simulation grid showing same pipeline produces visibly-distinctsignatures, 5 real governance catches during this build cataloged (QA caught 3 plugin-loader schema violations pre-demo; Brutus caught 3 contaminated blind-test prompts before 30 Opus calls fired; and more).
README excerpt
# Cognitive Signature IDE **Claude quietly syncs to your cognitive signature — voice (directives, conversational flow, idiomatic tells) + operational patterns (failure modes, reasoning chains, recurring decisions). The plugin injects your signature into Claude's context every prompt; Claude reads it and chooses whether to mirror it. Stronger enforcement (PreToolUse violation detector) ships in v2 — see *Architecture: Inject + Act*.** A Claude Code plugin that auto-promotes patterns from how you direct AI and from what you've learned through usage. Two functionalities across one pipeline. Three user types. Hands-off by default; user surfaced only on conflict. Built for the **Built with Opus 4.7** hackathon (Cerebral Valley + Anthropic), April 2026. --- ## The problem You open Claude Code. Every session starts cold. You spend the first 15 minutes bringing Claude up to speed on how you think. Then it forgets. Next session, you do it again. Every polished output is joint-authored — it came from you + Claude iterating — but nothing carries over. Your directing style, your reframes, your trust signals, your idiomatic tells evaporate at the end of every conversation. Worse: the failure modes you hit repeatedly, the reasoning shapes you converge on, the recurring decisions — none of those stack either. Every session, cold. **The data to fix this is already on your disk.** Claude Code stores every session as a JSONL. Every message you typed is there, tagged `type: user`. We cons