Skip to content

dnh33/prompt-cortex

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

37 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

prompt-cortex

                                          ╭─────────────────────────────╮
    ┌──────────────────┐                  │                             │
    │  "review my code" ├────────────────►│   ░░░░░░░░░░░░░░░░░░░░░░   │
    └──────────────────┘                  │   ░  prompt-cortex      ░   │
           what you type                  │   ░  350 templates      ░   │
                                          │   ░  622 index keys     ░   │
    ┌──────────────────┐                  │   ░  <2s matching       ░   │
    │  senior engineer  │                 │   ░░░░░░░░░░░░░░░░░░░░░░   │
    │  code review with │◄────────────────│                             │
    │  5-point checklist│                 ╰─────────────────────────────╯
    └──────────────────┘
           what claude gets                      you see nothing

The invisible layer beneath every prompt. 350 proven templates. Pure bash + jq. Zero latency. You type naturally — Claude receives expert-level framing.

prompt-cortex is what CLAUDE.md would be if it were alive.


Built for Superpowers

prompt-cortex is the invisible half of the superpowers stack. Two plugins, two layers, one pipeline:

    "review my code"
           │
     ══════╪══════════════════════════════════════════
     ░░░░░░▼░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░
     ░  PROMPT-CORTEX              invisible layer   ░
     ░  UserPromptSubmit hook      fires BEFORE      ░
     ░  ─────────────────────────────────────────     ░
     ░  + senior engineer code review template       ░
     ░  + 5-point checklist (correctness, security   ░
     ░    error handling, performance, naming)        ░
     ░  + injected as additionalContext               ░
     ░░░░░░│░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░
     ──────┼──────────────────────────────────────────
     ██████▼██████████████████████████████████████████
     █  SUPERPOWERS                 visible layer    █
     █  Skill auto-detection        fires DURING     █
     █  ─────────────────────────────────────────     █
     █  + requesting-code-review workflow             █
     █  + diff analysis, structured output            █
     █  + verification-before-completion              █
     ██████│██████████████████████████████████████████
     ══════╪══════════════════════════════════════════
           ▼
    Claude responds with expert framing
    AND a structured workflow

They complement, not compete. Cortex injects the what to think about. Superpowers provides the how to execute. Different lifecycle points, natural synergy. Every template declares which superpowers skills it pairs with via compatible_with — 11 of 14 superpowers skills are mapped.

Install both for the full stack:

claude plugin add github:dnh33/prompt-cortex
# + superpowers from marketplace

What it does

When you type a prompt, prompt-cortex analyzes it in the background and — if it matches a known pattern — silently injects a proven prompt template as additional context. You don't see it, but Claude does. The result: better responses without changing how you work.

Example: You type review my code. prompt-cortex detects a code review intent, injects a senior engineer code review template with a 5-point checklist (correctness, error handling, naming, performance, security), and Claude responds with structured, thorough feedback — all without you having to write a detailed prompt.

How it works

  1. SessionStart hook fires when Claude Code starts. Validates the template index, detects your project's language/framework, and confirms cortex is active.
  2. UserPromptSubmit hook fires on every prompt. A pure jq matching engine scores your prompt against 50 gold-tier templates in <2 seconds.
  3. If confidence is high enough (≥0.70), the best template is injected as additionalContext — invisible to you, visible to Claude.
  4. If confidence is medium (0.40-0.69), a hint is injected instead.
  5. If the prompt looks like it should be left alone (shell commands, continuations, already-structured prompts), cortex stays silent.

Full pipeline example

Prompt: "review my code"

 ┌─────────────────────────────────────────────────────────────────┐
 │ 1. Leave-it-alone detector                                     │
 │    13 signals checked → score 0.0 → passes through             │
 ├─────────────────────────────────────────────────────────────────┤
 │ 2. Preprocessor (v1.3)                                         │
 │    No pattern matched — direct action word, no preprocessing   │
 ├─────────────────────────────────────────────────────────────────┤
 │ 3. Keyword lookup                                              │
 │    "review", "my", "code" hit inverted index                   │
 │    → coding-001 "Code Review" is top candidate                 │
 ├─────────────────────────────────────────────────────────────────┤
 │ 4. Scoring                                                     │
 │    action  0.45  (exact match: review)                         │
 │    object  0.35  (match: code)                                 │
 │    signal  0.10  (intent signal hit)                           │
 │    keyword 0.02  (index key hit)                               │
 │    total   0.92  confidence                                    │
 ├─────────────────────────────────────────────────────────────────┤
 │ 5. Context filter                                              │
 │    No language/framework restrictions → passes                 │
 ├─────────────────────────────────────────────────────────────────┤
 │ 6. Result                                                      │
 │    action: "inject" at 0.92                                    │
 │    coding-001 body → additionalContext                         │
 └─────────────────────────────────────────────────────────────────┘

Conversational NLP (v1.3)

Before v1.3, matching relied on exact keyword hits. Natural phrasing — filler words, implicit actions, pronouns — degraded scores or caused misses. The v1.3 preprocessor strips noise and infers intent from conversational patterns before scoring begins.

Prompt Before v1.3 After v1.3
"something is wrong with the API" skip (no action match) defer — preprocessor infers debug, boosts candidates (conf 0.35 skip, but preprocessor extracts intent)
"make it faster" skip ("make" suppressed as shell cmd) defer — preprocessor maps "faster" → optimize via adjective_actions (conf 0.35 skip, intent extracted)
"how does the auth flow work" skip (conceptual question suppressed) defer — preprocessor extracts explain + subject "auth flow" (conf 0.40)
"can you help me debug this" defer (weak prefix match) defer — "can you help me" stripped, clean debug term boosts match (conf 0.47)

Install

# Clone the repo
git clone https://github.com/dnh33/prompt-cortex.git

# Install as Claude Code plugin
claude plugin add /path/to/prompt-cortex

Or install directly from GitHub:

claude plugin add github:dnh33/prompt-cortex

Escape mechanisms

prompt-cortex is designed to stay out of your way. If you ever want to bypass it:

Method How Scope
--raw prefix Start any prompt with --raw Single prompt
/cx off Type /cx off as a prompt Rest of session
/cx on Type /cx on to re-enable Rest of session

The engine also automatically stays silent for:

  • Shell commands (git status, npm install, etc.)
  • Slash commands (/commit, /help, etc.)
  • Continuations (ok, yes, looks good, ship it)
  • Already-structured prompts (XML tags, numbered lists, role assignments)
  • Conceptual questions (what is React?, how does X work?)

Superpowers skill coverage

Every template maps to 1-3 superpowers skills via the compatible_with field. See docs/superpowers-skills-reference.md for the full mapping and guidelines for new templates.

Skill coverage matrix
Superpowers Skill Templates that complement it
brainstorming API design, architecture, components, agents, RAG, embeddings, tool use, prompts, search, auth
test-driven-development Write tests, create functions, fix bugs, error handling, regex, concurrency, evals, guardrails, parsers
verification-before-completion Code review, performance, security audit, config, accessibility, deployment, cost optimization
systematic-debugging Debug error, fix bug, error handling, concurrency, type safety, logging
writing-plans API design, database, architecture, deployment, auth, caching, search, RAG, agents, embeddings
requesting-code-review Code review, refactor, PR review, docs, security audit, architecture, accessibility
writing-skills Prompt engineering, system prompts, chain-of-thought, few-shot, tool use, conversation design
dispatching-parallel-agents Agent design, batch processing
finishing-a-development-branch PR review, git workflow
using-git-worktrees Git workflow, deployment
receiving-code-review Refactor code, type safety

Commands

Command Description
/cortex:debug Show what cortex matched on your last prompt
/cortex:list List all available templates by category
/cortex:sync Rebuild the template index after changes

Templates

50 gold-tier templates across two categories:

  • coding (28 templates): Code review, debugging, testing, refactoring, API design, security audit, performance optimization, error handling, and more.
  • ai-workflows (22 templates): Agent design, RAG pipelines, prompt engineering, evaluation, embeddings, tool use, guardrails, streaming, cost optimization, and more.

Each template includes:

  • Senior expert role framing
  • Systematic approach (numbered steps)
  • Conditional guidance ("If no code provided, ask...")
  • Triggers and intent signals for accurate matching
  • Negative signals to avoid false positives

Adding templates

See CONTRIBUTING.md for the template schema and guidelines.

Quick start:

  1. Create a .md file in data/prompts/<category>/
  2. Add YAML frontmatter with required fields
  3. Validate: bash scripts/validate-template.sh your-template.md
  4. Rebuild index: /cortex:sync or bash scripts/build-index.sh

Architecture

prompt-cortex/
  .claude-plugin/plugin.json       # Plugin manifest
  hooks/
    hooks.json                     # Hook registrations
    run-hook.cmd                   # Cross-platform wrapper
    cortex-match                   # UserPromptSubmit hook
    cortex-session-init            # SessionStart hook
  data/
    intents.json                   # Action/object taxonomy
    index.json                     # Pre-built inverted index
    prompts/                       # Template library
      coding/*.md
      ai-workflows/*.md
  scripts/
    match.jq                      # Core matching engine
    build-index.sh                 # Index compiler
    validate-template.sh           # Template validator
  tests/
    run-tests.sh                   # Integration tests
  commands/
    debug.md                       # /cortex:debug
    list.md                        # /cortex:list
    sync.md                        # /cortex:sync

Tech stack: Pure bash + jq. No Python, Node, or other runtime dependencies. Cross-platform via .cmd polyglot wrapper (Windows/macOS/Linux).

Troubleshooting

Hooks not firing

If prompt-cortex doesn't seem active after installation, the plugin hooks may not be loading (known Claude Code issue #10225). Workaround: copy the hook configuration to your user settings:

# Copy hooks to user settings
cat prompt-cortex/hooks/hooks.json
# Then add the hooks to ~/.claude/settings.json under the "hooks" key

Verifying it works

  1. Run /cortex:debug after sending a prompt
  2. Check .cortex/usage.jsonl in your project directory for telemetry
  3. Run the test suite: bash tests/run-tests.sh

Index out of date

If you've added or modified templates, rebuild the index:

bash scripts/build-index.sh
# or use the command:
/cortex:sync

Local telemetry

prompt-cortex logs match results to .cortex/usage.jsonl in your project directory. This is local-only, contains no prompt content (only hashes), and is gitignored. Use it to understand what's being matched and at what confidence.

License

MIT — Daniel Hjermitslev (@dnh33)

About

Invisible prompt intelligence layer for Claude Code. 350 expert prompts that inject themselves into your workflow.

Resources

License

Contributing

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors