| provider | claude |
| name | context-priming |
| version | 1.1.8 |
| updated | "2026-03-09T00:00:00.000Z" |
| description | Dynamic context provider for codebase knowledge. This skill should be used when working on code, answering questions about architecture, or making changes that require understanding of project structure. Maintains area-specific reference files that are loaded on demand. Self-updating based on observed drift between references and actual code. Load this skill early in work sessions - it contains essential workflow and update protocols. |
Context Priming
Provide dynamic, accurate context about this project's architecture and code. Load knowledge on demand, update when drift is observed, self-improve based on experience.
When to Invoke
Load context when:
- Starting work requiring codebase knowledge not in current conversation
- Making changes to unfamiliar areas
- Answering questions about project architecture or patterns
- User asks about functionality and current context is insufficient
Do NOT load when:
- Information is clearly present from last 3-5 exchanges
- Task is trivial and doesn't require architectural understanding
- Already loaded relevant areas in this session (unless significant time/work has passed)
How to Use
- Load area-index.md first - understand available areas, their update dates, load-when triggers
- Load relevant area(s) - based on current task and index guidance
- Follow area's "when to expand" - each area indicates when to open specific files
- Evaluate freshness - compare area's
updated date against today (from system prompt); consider git activity if concerned
Callouts
Prefix operational updates with Priming: to signal skill activity. Keep brief - one line.
When to call out:
- Loading an area:
Priming: Loading backend - API work detected.
- Loading multiple:
Priming: This spans backend and database areas, loading both.
- Staleness concern:
Priming: backend looks stale (Nov 20), will verify as I go.
- Making a note:
Priming: Quick note added - auth order matters.
- Skipping load:
Priming: Already have frontend context from earlier.
Don't call out: Every file read, minor internal decisions, routine checks.
Update Behavior
| Situation | Action |
|---|
| Loaded info contradicts code in misleading way | Update reference in stride |
| Minor drift, no practical impact | Let it slide |
| Unsure if update needed | Note concern, ask at end of task |
| Multiple areas need restructuring | Consult user before proceeding |
| New feature added affecting an area | Update area reference |
| User says "context priming needs to know this" | Capture in relevant area |
| User says "you should have known this" | Analyze failure, propose fix (see below) |
Small updates: Make concisely, in stride, no announcement needed.
Large restructuring/defrag: Consult first. See references/maintenance.md.
"You Should Have Known" Response
When user indicates priming failed:
-
Identify cause: Which of these?
- Area not loaded when it should have been (bad load-when criteria)
- Area loaded but info missing (incomplete reference)
- Info existed but outdated (stale reference)
- Area doesn't exist (missing area)
- Judgment call to skip loading was wrong (bad heuristic)
-
Propose fix based on cause:
- Content/index updates → do in stride
- SKILL.md or maintenance.md changes → explain problem, propose change, await approval
-
Surface immediately - don't wait for task end. User will indicate if disruptive.
Permission Model
| File | Permission |
|---|
| references/areas/*.md | Update in stride |
| references/area-index.md | Update in stride |
| Add/remove/split areas | Suggest, light confirmation |
| SKILL.md | Explain, propose, await approval |
| references/maintenance.md | Explain, propose, await approval |
Git Usage
Use git diff or git log sparingly:
- Before major refactors to verify scope
- When user questions accuracy of loaded info
- To check activity since area's last update date when staleness suspected
Not for routine small changes.
Context Compaction
No direct visibility into remaining context space. When user flags compaction is imminent, or after long exploration sessions: consider if valuable learnings should be captured before context is summarized.
Self-Improvement
Area notes (in area-index.md) capture learnings:
- Up to 10 notes per area, quality over quantity
- Format: actionable guidance ("when X, do Y" or "don't assume Z")
- Remove notes that no longer apply
- If skill behavior is flawed, suggest improvement to user
If suggestion timing/frequency becomes disruptive, user will indicate - adjust accordingly.
Operational Principles
- Concise updates: prefer short phrase tweaks over paragraph additions
- Don't bloat references - information density is the goal
- When on fence about loading, lean toward loading
- Trust the index's load-when triggers; refine them when wrong
- No hardcoded "status" - freshness is evaluated dynamically
Skill Location & Cross-Provider Note
Canonical path (relative to project root):
.claude/skills/context-priming/
├── SKILL.md # This file
├── references/
│ ├── area-index.md # Index of all areas (load first)
│ ├── maintenance.md # Defrag, pruning, area separation doctrine
│ └── areas/
│ └── [name].md # Deep reference per area
Resolution order — try these paths in order, use the first that exists:
- Primary:
<project-root>/.claude/skills/context-priming/ — the canonical, most up-to-date copy. Always check here first.
- Fallback: The directory containing this SKILL.md (i.e. your own skill folder). If there is no
.claude/ directory in the project, the references were likely copied alongside this file.
This matters when the skill is deployed to another provider (Codex at .agents/skills/, Gemini at .gemini/skills/). Those folders may contain just this SKILL.md for discovery, with the area files living in .claude/. But if the project has no .claude/ directory at all (e.g. a pure Codex or Gemini project), fall back to reading from your own skill directory.
Resolved paths (primary):
- Area index:
<project-root>/.claude/skills/context-priming/references/area-index.md
- Area files:
<project-root>/.claude/skills/context-priming/references/areas/<name>.md
- Maintenance:
<project-root>/.claude/skills/context-priming/references/maintenance.md
Resolved paths (fallback — relative to this SKILL.md):
- Area index:
./references/area-index.md
- Area files:
./references/areas/<name>.md
- Maintenance:
./references/maintenance.md
Files
references/area-index.md - compact index of all areas
references/maintenance.md - defrag, pruning, area separation doctrine
references/areas/[name].md - deep reference per area