context-os
>-
Works with
Claude CodeCursorCodex CLIGitHub CopilotGemini CLI
--- name: context-os description: >- license: MIT --- # ContextOS Persistent project intelligence for AI coding agents. This skill is a **router**: it tells you *when* to act and *which reference to read* for the details. Keep this file in context; load a reference only when its situation applies. ## Prime directive **Conversations are temporary. The repository is the durable memory.** Persist anything that would be expensive to reconstruct into `.context/`. Never depend on chat history or model memory for project truth. ## When to engage (thresholds) Engage ContextOS for **substantial** work: a new project or product, a non-trivial feature, a multi-step change, or any work you will hand off or resume later. Do **not** engage for trivial requests — typo fixes, renames, spacing, a one-line change. Do not create `.context/` or a PRD for these. ContextOS should feel proportionate, never bureaucratic. When unsure, prefer a light touch and scale up only if the work grows. ## Routing table | Situation | Do this | Reference | | :-------- | :------ | :-------- | | Substantial **new project** | Understand intent → improve request → PRD → architecture → plan → initial STATUS | [initialization](references/initialization.md) | | **Vague / sprawling request** | Compile it into an execution spec before coding | [prompt-compiler](references/prompt-compiler.md) | | Need a **PRD** (create or evolve) | Generate/update structured requirements with IDs | [prd-engine](references/prd-engine.md) | | Working with **memory files** | Read/write `.context/` correctly | [project-memory](references/project-memory.md) | | Deciding **how much to load** | Apply the L0–L3 budget; load least-sufficient | [context-budget](references/context-budget.md) | | Finished meaningful work | Update `STATUS.md` at the boundary | [status-protocol](references/status-protocol.md) | | Made an **architectural decision** | Record an ADR entry in `DECISIONS.md` | [decision-memory](references/decision-memory.md) | | **Resuming**, or memory looks stale | Rehydrate and reconcile against reality | [context-recovery](references/context-recovery.md) | | **Ending a session** / limit near | Write/refresh `HANDOFF.md` | [handoff-protocol](references/handoff-protocol.md) | | **Switching model/agent** | Save state, hand off, resume elsewhere | [model-switching](references/model-switching.md) | | Delegating to a **sub-agent** | Build a task packet; expect a result summary | [multi-agent-context](references/multi-agent-context.md) | | Signs of **drift** | Stop, verify against code, repair memory | [context-drift](references/context-drift.md) | | Before/after **compaction** | Snapshot first; recover cleanly after | [context-compaction](references/context-compaction.md) | ## Automatic lifecycle ``` NEW LARGE PROJECT → initialize .context/ VAGUE LARGE REQUEST → compile the prompt NEW PRODUCT → PRD → architecture → plan → initial STATUS SESSION START → restore: INDEX → STATUS → HANDOFF → verify git TASK COMPLETE → update STATUS NEW DECISION → record in DECISIONS BEFORE COMPACTION → snapshot STATUS / DECISIONS / HANDOFF SESSION END → write HANDOFF NEW AGENT → rehydrate from .context/ ``` The user should not need special commands for any of this. Optional `/context-*` commands and a `contextos` CLI exist for convenience but are never required. ## Core operating rules 1. **Smallest sufficient context.** Load the least high-signal context needed to make the next decision correctly. More context is not better context. Start from `INDEX.md` → `STATUS.md`, then open only what the task needs. 2. **Repository reality wins.** When sources disagree, trust: current user instruction → actual code/tests → accepted decisions → PRD → architecture → status → handoff → archive. If a doc is stale, update it; don't blindly follow it. 3. **Persist at boundaries, not constantly.** Update `STATUS`/`DECISIONS`/`HANDOFF` at meaningful moments (phase/feature done, new blocker, decision, test run, commit, handoff, before compaction) — not after every small edit. 4. **Never fake validation.** Only report a check as passing if you ran it. Record real results in `STATUS.md`. 5. **Never persist secrets.** No API keys, tokens, passwords, or private data in `.context/`. See [SECURITY](../../SECURITY.md). 6. **Don't reverse decisions silently.** Read `DECISIONS.md` before architectural changes; supersede decisions explicitly, never delete them. 7. **Preserve intent.** When compiling prompts or writing PRDs, improve clarity and structure but do not invent product decisions the user didn't make. Ask only genuinely blocking questions. ## Getting `.context/` into a project Scaffold the `.context/` files in any of these ways: - run `npx github:pokhrelboss/context-os init .` (no clone needed), or `contextos init` if the CLI is on PATH; - or create the files directly following [project-memory](references/project-memory.md) — the references fully specify every file, so no templates or network are required. Existing project? Start with repository discovery — see [context-budget](references/context-budget.md) and [context-recovery](references/context-recovery.md) — before writing anything. ## Reference index - [initialization](references/initialization.md) — new-project bootstrap - [prd-engine](references/prd-engine.md) — PRD generation & evolution - [prompt-compiler](references/prompt-compiler.md) — vague → execution spec - [project-memory](references/project-memory.md) — the `.context/` files - [context-budget](references/context-budget.md) — L0–L3 & discovery - [status-protocol](references/status-protocol.md) — STATUS updates - [decision-memory](references/decision-memory.md) — ADR entries - [context-recovery](references/context-recovery.md) — rehydrate & reconcile - [handoff-protocol](references/handoff-protocol.md) — HANDOFF - [model-switching](references/model-switching.md) — Claude ⇄ Codex ⇄ others - [multi-agent-context](references/multi-agent-context.md) — task packets - [context-drift](references/context-drift.md) — detect & repair drift - [context-compaction](references/context-compaction.md) — survive compaction
More AI & ML skills
writing-shape
mattpocock/skills
Writing, exploit: shape raw material into an article, paragraph by paragraph.
279.3k
writing-fragments
mattpocock/skills
Writing, explore: mine raw fragments, no structure yet.
279.2k
full-output-enforcement
leonxlnx/taste-skill
Overrides default LLM truncation behavior. Enforces complete code generation, bans placeholder patterns, and handles token-limit splits cleanly. Apply to any task requiring exhaustive, unabridged output.
275.4k

