microsoft-learn

CCF — Claude Context First

English · Tiếng Việt · 简体中文

A plugin for Claude Code that keeps your AI coding assistant disciplined: plan first, keep the project spec fresh, check work against the spec, and run independent tasks in parallel.

Why use it

Plain Claude Code is a sharp assistant with a short memory. Over a long session it forgets your project rules, lets planning slide into editing files, and never reminds you to update the docs. CCF adds three habits that stick:

• Plan before code. /ccf:plan only runs in plan mode, so planning stays read-only and reviewable. No accidental edits.
• Spec stays fresh. Hooks (small scripts that fire on session events) nudge you to update the spec when the code changes. Lessons you learn get saved into system memory, so mistakes stop repeating.
• Check before done. /ccf:check reviews each finished task against the spec with two independent reviewers. Nothing counts as done until it passes.

Install

In Claude Code:

/plugin marketplace add naniiluja/ccf
/plugin install ccf@ccf

Or in one step: npx @naniiluja/ccf

Then open Claude Code in your project folder and run /ccf:init.

The 5 commands

Typical flow: /ccf:init → /ccf:plan → /ccf:cook → /ccf:check → /ccf:updatespec.

Under the hood

You don't need this section to use CCF. It's here for the curious.

• Hooks are the deterministic layer. Commands and agents are prompts, and a model can ignore a prompt. Hooks are scripts that run on session events no matter what: they block /ccf:plan outside plan mode, nudge a spec update when code changed, and re-load your in-progress task after a compact.
• Agents are 6 read-only helpers: one reads slices of your codebase, one fetches best practices from official docs, one drafts spec text, two review your work, and one tries to disprove each blocking finding before you act on it. None of them write code. You write the code, directly in the session.
• Scripts are 13 command-line tools that only run when you (or a command, after asking) start them, because they change files or call an outside API. Among them: plan-lint.mjs shows, before a run, which tasks would be forced into waves of one because they edit the same shared file, prune-archive.mjs clears out task files of old finished iterations (the archive and git history keep the record), and memory-audit.mjs tells /ccf:updatespec when the system memory has grown enough to need a cleanup, spec-budget.mjs measures how many bytes of spec every session pays for, and eval-changelog.mjs writes each release's eval scores (or why none ran) into CHANGELOG.md.
• Docs lookup built in. The plugin ships Context7 and Microsoft Learn (MCP servers), so design advice cites real documentation instead of model memory.

Full internals reference: plugins/ccf/README.md. Requires Node ≥ 18 for the hooks.

License

MIT

Acknowledgements

First released in the LINUX DO community. Thanks for the support and feedback.