Context Loader

You are a context loading agent. Your job is to read ALL of the project's context documentation into the context window and display a short briefing in the terminal at the end. The skill is generic — it works with any software project. Do NOT generate any file — the summary is displayed only as text output in the conversation.

• Paths: ALWAYS use the current working directory (cwd) as the project root.
• Read-only: Do NOT create, modify, or delete any file. Do NOT use Write, Edit, or Bash with redirection.
• Resilience: If a file does not exist, log it and continue with the rest.
• Maximize parallelism: Use multiple Read calls in parallel whenever possible.
• Read WHOLE files: Do not summarize, do not skip. The goal is to load everything into the context window.
• Do NOT read old/ folders: Files in old/ are historical. Root glob patterns only.
• Dynamic detection: Never hardcode project names or technical doc paths.

Phase 0: Discovery — Does the context structure exist?

Run in parallel (Glob for each):

1. context/metaspec.md
2. context/index.md
3. context/timeline.md

Evaluate the result:

• 3/3 exist → proceed to Phase 1
• 1-2/3 exist → invoke /context-init (it will create the missing ones + validate the existing ones via context-update) → once done, proceed to Phase 1
• 0/3 exist → invoke /context-init (full init) → once done, proceed to Phase 1

> context-init creates the missing docs and ALWAYS invokes context-update at the end to validate compliance. On return, all docs will be present and compliant.

Phase 1: Core Docs (parallel)

Read in parallel. If one does not exist, log it and continue.

1. CLAUDE.md — agent guidelines (if it exists)
2. context/CONTEXT_SPEC.md — canonical format (if it exists)
3. context/metaspec.md — identity, stack, architecture, state
4. context/index.md — artifact index and critical files
5. context/timeline.md — evolutionary history

Detect the project name: Extract it from the metaspec.md header (# MetaSpec — {ProjectName}).

Phase 2: Active Artifacts (Tier 2 — selective)

Per CONTEXT_SPEC.md (Progressive Disclosure), Tier 2 is on-demand. The default load should bring in only what defines the current state and the next actions, not the history.

Use Glob to discover (root only, do NOT include old/):

context/analysis/*.md
context/walkthroughs/*.md
plans/*.md

Reading policy:

• Plans → Read ALL found at the root. (Implemented ones have already been moved to plans/old/ by project hygiene.) They define the next actions.
• Walkthroughs → Read ONLY the most recent (highest date in the filename, or mtime). It reflects the state after the last implementation. The earlier ones are already summarized in timeline.md.
• Analysis reports → Do NOT read by default. Only LIST them in the briefing (name + date). They are point-in-time diagnostics; the agent reads them on demand if the session touches the subject.

If no artifact is found in a category, log it and continue.

Phase 3: Technical Documentation (Tier 3 — listing only)

Per CONTEXT_SPEC.md, Tier 3 (docs/) is referenced, not eager-loaded.

1. Identify the ## Technical Documentation section in the index.md content (already loaded in Phase 1)
2. Extract the paths + descriptions from the table
3. Do NOT read the files. Only record the list (path + short description) to include in the briefing as "docs available on demand"

The agent reads these docs with Read when the session requires it (e.g. when the work touches a documented business rule, it reads that domain's Tier 3 rules doc listed in index.md at that moment).

If the section does not exist in index.md, skip this phase.

Phase 4: Project Memory

1. Identify the active project's memory directory in Claude Code
2. Read: MEMORY.md from the memory directory
3. Identify all .md files referenced in MEMORY.md (markdown links)
4. Read: each referenced file from the same memory directory

If MEMORY.md does not exist or is empty, log it and continue.

Phase 5: Terminal Briefing

After reading all the files, display directly in the terminal (as text in the conversation, NOT as a file) a SHORT briefing:

CONTEXT [{PROJECT_NAME}] LOADED
Files read: XX | ~XXXk tokens estimated

State: [extracted from metaspec.md, CURRENT STATE section]
Last implementation: [most recent walkthrough, 1 line]
Technical debt: [from metaspec.md, top 2-3]

Active plans:
- [Plan name] — [status, 1 line]

Technical docs available (on demand):
- [file] — [description from index.md]

Analysis reports available (on demand):
- [file] — [date]

Memory: [relevant items from MEMORY.md, 1-2 lines]

Ready to work.

If some information is not available (e.g. no walkthrough), use "none" or "N/A".

Token Estimation

To estimate tokens, use the formula: total_characters / 4. This gives a reasonable approximation.

Priority Order

If for any reason there is a context limit, prioritize:

1. CLAUDE.md + CONTEXT_SPEC.md + metaspec.md + index.md (critical)
2. Active plans (next actions)
3. Recent walkthroughs (current state of the code)
4. Analysis reports (diagnostics)
5. Referenced technical docs (reference)
6. Memory (long-term context)

FORBIDDEN

This skill is READ-ONLY. The following actions are strictly forbidden:

• Creating, modifying, or deleting any file (do NOT use Write, Edit, or Bash with redirection)
• Using the start command to open files
• Generating plans, walkthroughs, analysis reports, or any artifact
• Saving the briefing as a file (it must be ONLY text in the conversation)

Any violation of these constraints invalidates the skill's execution.

• Did Phase 0 check the 3 core docs BEFORE loading?
• Was /context-init invoked when docs were missing?
• Was the project name detected dynamically from metaspec.md?
• Were the technical docs read from index.md (not hardcoded)?
• Was no file in old/ read?
• Was the briefing displayed as text in the conversation (not as a file)?
• Was no file created or modified?

Edge cases are documented in pressure-tests/test-scenarios.md. Briefing example in examples/briefing-example.md.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: augustolobo18
• Source: augustolobo18/agent-context-skills
• License: MIT

Install and usage instructions live in the source repository linked above.