Skills

Six auto-discovered skills that teach the host AI to delegate entire tasks to SubQ's 12M-token context engine.

Design Philosophy

9 Python PreToolUse hooks intercepting every Read/Grep/Glob 6 declarative skills delegating entire tasks
Not interception—delegation. The model decides when SubQ's context window advantage matters.

The v1 design intercepted every tool call, polluting the UI with status messages and burning tokens on simple lookups. Skills invert the relationship: ~100 tokens per skill at startup (progressive disclosure), and Claude Code's own Read/Grep/Glob remain available for targeted operations. The host decides when to delegate.

Skill Catalog

Two skills auto-trigger on conceptual queries (and can be invoked manually); four are user-invoked for heavyweight operations.

AUTO / USER

/subq-search

LLM-powered codebase search—find code by meaning, not just text. Cross-file conceptual queries, blast radius analysis, identifier discovery.

subq context search | answer | blast-radius | identifiers

AUTO / USER

/subq-context-pack

Structural analysis—directory trees, function signatures, context packs, raw dumps. Understand architecture without reading every file.

subq context tree | skeleton | pack | dump

POST-LAUNCH

/subq-spec

Implementation specification generation with interactive interview. Produces structured specs grounded in codebase analysis.

subq spec "<description>"

POST-LAUNCH

/subq-review

Code review for uncommitted changes. Full-codebase context means review catches cross-file regressions.

subq review

POST-LAUNCH

/subq-audit

Codebase quality audit across complexity, testing coverage, and security. Produces actionable findings.

subq audit

POST-LAUNCH

/subq-bench

A/B token and cost comparison between SubQ and Claude Code on the same query. Parallel runner with timing.

parallel SubQ vs Claude Code

When to Delegate

If the query requires understanding relationships across files or reasoning about architecture, delegate. If it's a lookup, don't.

Query Type
Delegate to SubQ
Keep in Host
Cross-file conceptual
Yes
Blast radius analysis
Yes
Structural overview
Yes
Full codebase review
Yes
Single-file read
Yes
Symbol grep
Yes
Known-path lookup
Yes
Simple edits
Yes

Auto-Discovery

Skills are discovered automatically—no explicit manifest key needed. The flow from canonical source to provider-specific output:

Canonical TS Skill definitions live as TypeScript in the SubQ CLI monorepo (apps/cli/src/skills/definitions/).
Generator subq skills generate --provider claude|codex|gemini|all emits provider-specific SKILL.md files at build time.
SKILL.md Claude Code scans skills/*/SKILL.md at the plugin root. ~100 tokens loaded at startup; full instructions on invocation.
No vendor lock-in: Adding a new provider means writing a thin adapter and a SKILL.md template—not duplicating six skill bodies.