Search skills/

Continuous Claude Claude skills

A personal Claude Code infrastructure: ~160 skills, 32 agents and 30 hooks that inject skill hints, isolate subagent context to files, and fight compaction.

parcadei/Continuous-Claude-v3 160 skills Planning, TDD & debuggingMemory & context
Our verdict Recommended with caveats

An ambitious and genuinely instructive personal system with real file-based context-isolation patterns worth studying, but it is large, intertwined with hooks, a daemon, a database and a vendored tracing plugin, and shows no recent activity, so install only after reading what each hook does.

3/5 Depth expertise shipped ! Security Broad access 9mo Maintenance Updated 9 mo ago 3.9K Adoption stars OK License MIT, commercial OK

What Continuous Claude does

Continuous Claude treats compaction (losing nuance when Claude Code's context fills up) as the core problem to solve, under the stated mantra 'compound, don't compact.' A hook fires on every message and prints a skill-activation checklist (critical/recommended/suggested/optional skills and agents) that Claude is told to consult before responding. Underneath that are three mechanisms: YAML handoffs for token-efficient context transfer between sessions, a 'TLDR' 5-layer code analysis (AST, call graph, control flow graph, data flow graph, semantic index) that the README claims cuts token use by about 95% versus reading whole files, and a background daemon that extracts learnings and recalls them in later sessions.

agent-context-isolation and agent-orchestration are two of the clearest skills: they instruct Claude to spawn background subagents that write results to files rather than returning a full transcript via TaskOutput (said to dump 70k+ tokens into context), and to verify work with tests rather than by reading agent output. The repository is less a curated pack than an evolving personal system: skill folders include an archive/ subdirectory of retired approaches, a vendored Braintrust-tracing plugin with its own setup script and API key, a math system, a memory system backed by what the README calls a daemon, and database setup steps requiring pgvector and superuser access. The last commit in our dossier was from January 2026, several months before this review, so treat its currency and compatibility with the latest Claude Code as unverified.

Good for

  • Power users who want to study or borrow specific context-management patterns (file-based agent handoffs, token-efficient code summaries)
  • Long Claude Code sessions that repeatedly hit compaction
  • Readers interested in the author's architecture writeup rather than a drop-in install

Not for

  • Anyone wanting a small, easily audited skill pack
  • Users who want an actively maintained, currently-tested system (no recent commits in our dossier)

What's inside depth 3/5

160 skill folders averaging about 4 KB (several in an archive subfolder of retired approaches), backed by 30 hooks and 32 agents per the README, but most sampled skills are short rule files rather than deep references or scripts.

Skills
160
Reference docs
7
Scripts
0
Tests / evals
0
Templates
0
Avg SKILL.md
4.1 KB

Counted from the repository file tree. How we score depth.

Dependencies and costs

  • Claude Code CLI
  • A database with the pgvector extension (requires superuser or rds_superuser) for the memory system
  • jq for the vendored Braintrust-tracing setup script
  • Optional Braintrust API key for trace observability
  • Optional Perplexity and NIA integrations (stated as optional, not required)
  • Telemetry: The vendored braintrust-tracing plugin sends full Claude Code session data (every tool call, file read, edit and terminal command) to Braintrust once configured with an API key; this is opt-in per the plugin's own setup script, not on by default for the rest of the system.

Security review

Broad access

This skill asks for more access than most, or sends data to another service. That is usually a design choice rather than bad intent, but you should know before you install it.

  • Pre-approves broad shell commandsMany skills declare frontmatter such as allowed-tools: [Bash, Task, Read, Glob, Grep, Write] or [Bash, Read], which pre-approves unrestricted Bash (13 SKILL.md files list Bash)..claude/skills/explore/SKILL.md:4
  • Turns off permission prompts or safety checksThe cli-reference skill tells the agent that headless claude -p spawns need --dangerously-skip-permissions ('Headless requires this') and gives it in the standard spawn pattern and a CI example..claude/skills/cli-reference/SKILL.md:165

A heavyweight Claude Code environment: a wizard replaces your ~/.claude with 109 skills, 32 agents and 30 hooks that run on every session start, prompt, tool call and session end, backed by a local Postgres memory and a background daemon that mines past session transcripts. Many skills pre-approve unrestricted Bash and one teaches headless runs with permission prompts disabled. An optional Braintrust tracing hook sends full session content to a third party once you add its API key. The hook and wizard code was not available to us.

Worth knowing

  • Changes files outside your projectThe setup wizard moves your existing ~/.claude to an archive, copies 32 agents, 109 skills, 30 hooks and rules into ~/.claude, edits settings.json, and starts a Docker PostgreSQL stack; a background daemon later spawns headless Claude to extract learnings from past sessions.README.md:202
  • Installs packages or toolsWizard installs Python packages (sympy, z3-solver, pint, shapely), Docker services, the TLDR tool and an optional Lean toolchain.README.md:946
  • Installs with curl | shOptional Lean setup runs curl https://raw.githubusercontent.com/leanprover/elan/master/elan-init.sh -sSf | sh (third-party installer).README.md:949
  • Makes network requestsThe vendored braintrust-tracing plugin and braintrust-tracing skill hook SessionStart/PostToolUse/Stop/SessionEnd and send every tool call and turn to api.braintrust.dev once BRAINTRUST_API_KEY is configured; other skills call Firecrawl, Tavily and GitHub MCP servers with your keys..claude/plugins/braintrust-tracing/skills/trace-claude-code/SKILL.md:36
  • Needs an API key for its own serviceBRAINTRUST_API_KEY, FIRECRAWL_API_KEY and GITHUB_PERSONAL_ACCESS_TOKEN are read from ~/.claude/.env or mcp_config.json and used only with their own services..claude/skills/braintrust-analyze/SKILL.md:105
  • Relies on a CLI or package we could not reviewSetup and runtime depend on uv, Docker, the tldr tool, pyright/ruff, Claude Code's claude -p and a Postgres+pgvector database whose code is not among the files we have.README.md:174

What we checked

The repository (943 files) on Oct 29, 2026. Not a full review: 898 of 943 files were not downloaded (large repository; we reviewed the skill folders); it relies on code outside the repository: 898 readable files in the repo were not available to us, including all 30 hook implementations, the setup wizard and uninstall code, the memory daemon and the agents; only README.md and the SKILL.md files were reviewed; PyPI packages installed by the wizard (sympy, z3-solver, pint, shapely, etc.) and the TLDR tool; Braintrust, Firecrawl, Tavily and GitHub MCP services and the vendored braintrust-tracing plugin's setup script.

NVIDIA SkillSpector static scan: 60 pattern hits (Agent Snooping 22, Excessive Agency 11, Tool Misuse 6, Data Exfiltration 6, Prompt Injection 4, Privilege Escalation 4). The 'Agent Config Directory Access' hits are debugging commands that read ~/.claude logs and state files; the 'External Model or Provider Selection' hits are documented claude -p invocations in cli-reference; the 'credential access' hits are .env mentions for Braintrust keys; rm hits are troubleshooting and uninstall instructions.

A review, not a certification. Skills change after we check them — read the files before installing.

License

MIT — commercial use allowed.

Can it run unattended?

Not really. It is a context-management layer for interactive sessions with a per-message activation check; it is not designed around a verification loop for unattended runs.

Alternatives to Continuous Claude

Planning with Files

OthmanAdi/planning-with-files

Keeps task_plan.md, findings.md and progress.md on disk and re-injects them via hooks, so a plan survives /clear, compaction and crashes.

Recommended Broad access Depth 4/5

Superpowers

obra/superpowers

A 15-skill development methodology that makes coding agents clarify the goal, write a plan, work through it with subagents and verify before calling it done.

Recommended Check first Depth 3/5

Context Mode

mksglu/context-mode

MCP server plus hooks that keep bulky tool output out of the context window, index it in local SQLite, and restore session state after compaction.

With caveats Broad access Depth 3/5

A very large Claude Code harness: 311 skills, 71 agents, commands, hooks, memory, rules and a scanner, with adapters for Codex, Cursor and others.

With caveats Broad access Depth 3/5

FAQ

What is Continuous Claude?

It is a personal Claude Code infrastructure project combining roughly 160 skills, 32 agents and 30 hooks to fight context compaction: it injects skill-activation hints per message, isolates subagent output to files, summarizes code with a 5-layer token-efficient analysis, and extracts learnings across sessions.

How do I install it?

The repository is meant to be cloned or adapted into your own Claude Code project's .claude directory; it is not distributed as a simple plugin install, and parts of it (the memory system) require setting up a pgvector-enabled database.

Is it actively maintained?

Our dossier's last recorded push was in January 2026, so verify the hooks and skills still work with your current Claude Code version before relying on it.

Does any part of it send my data anywhere?

The core system runs locally, but the vendored braintrust-tracing plugin sends full session traces (tool calls, edits, conversation turns) to Braintrust once you configure it with an API key; it is opt-in and documented, not enabled by default.

Sources: GitHub repository (stars, license, file tree, last push Jan 26, 2026). Review based on the README and SKILL.md files as published.