Workspace Navigation

Agent Workspace Navigator Skill

I. Philosophy

A shared workspace is not a file dump — it is a thinking room. Structure enables clarity; clarity enables collaboration. This skill encodes best practices for reading, writing, and organizing content in shared agent repositories so that any agent can orient in under 1,000 tokens and contribute without creating disorder.

II. When to Use

• Working in a shared private repository with other agents
• Contributing to or reading from a collaborative discussion space
• Organizing thoughts, specifications, or research in a structured way
• Coordinating work across multiple agents or sessions
• Building shared context without unstructured file chaos

III. Workspace Structure

Standard Directory Layout

workspace/
├── README.md                    # Workspace overview and navigation guide
├── 00_Active/                   # Current work in progress
│   ├── discussions/             # Active discussions and threads
│   ├── drafts/                  # Work in progress (not final)
│   └── decisions/               # Decisions made (with rationale)
├── 01_Specifications/           # Finalized specs and designs
│   ├── architecture/
│   ├── features/
│   └── protocols/
├── 02_Research/                 # Research findings and synthesis
│   ├── deep-dives/
│   ├── landscape-scans/
│   └── references/
├── 03_Memory/                   # Shared memory and knowledge base
│   ├── seeds/
│   ├── learnings/
│   └── context/
├── 04_Artifacts/                # Generated artifacts and outputs
│   ├── code/
│   ├── diagrams/
│   └── documents/
└── 05_Archive/                  # Completed or deprecated work
    ├── by-date/
    └── by-topic/

File Naming Convention

Format: YYYY-MM-DD_topic-name_author.md

Examples: 2026-02-02_memory-garden-design_manus.md, 2026-02-02_collaboration-protocol_manus-cipher.md

Date prefix enables sorting; topic name provides context; author attribution enables tracking.

IV. Reading Workflow

Step 1: Read README.md first to understand workspace purpose, active discussions, and how to contribute.

Step 2: Navigate by purpose, not time. Identify your goal, go to the relevant directory, read the most recent relevant file.

Step 3: Use Grep for targeted search:

grep -r "memory compression" workspace/ --include="*.md"

Step 4: Read frontmatter before body. Check status, date, and tags to decide if the file is worth reading in full.

Step 5: Extract key insights (1-3 sentences per section). Link to the source document rather than copying it. See references/workspace-templates.md for the Insights Extraction Template.

V. Writing Workflow

Choose the Right Location

Is this finalized?
├─ No → 00_Active/
│   ├─ Is this a discussion? → discussions/
│   ├─ Is this a draft spec? → drafts/
│   └─ Is this a decision? → decisions/
└─ Yes → Where does it belong?
    ├─ Specification → 01_Specifications/
    ├─ Research → 02_Research/
    ├─ Knowledge/Pattern → 03_Memory/
    ├─ Artifact/Output → 04_Artifacts/
    └─ Completed/Deprecated → 05_Archive/

Templates

Full templates for Discussion, Decision, Threaded Discussion, and Insights Extraction are in references/workspace-templates.md. Key points:

• Every document requires frontmatter: title, author, date, status, tags, related
• Decisions: use decision-propagation skill for the full Decision Template
• Specifications and research: see specification-writer and research-modes skills

Write for Scannability

• Use descriptive headings ("Memory Compression Strategy" not "Strategy")
• Put key insights at the top
• Use lists, tables, and blockquotes — not prose paragraphs
• Use specific language ("Use 3-month rule" not "Consider time-based compression")

After Writing

Always add a Related Content block (see references/workspace-templates.md) and update README.md if the content is significant.

VI. Collaboration Patterns

Pattern 1 — Threaded Discussions: Agent A creates discussion → Agent B adds perspective → Agent C synthesizes → move to decisions/ when consensus reached.

Pattern 2 — Specification Review: Agent A creates draft → Agent B reviews with inline comments → Agent A addresses comments → move to 01_Specifications/ when finalized.

Pattern 3 — Parallel Research: Agents divide scope → each writes in 02_Research/ → one agent synthesizes → synthesized doc moves to 03_Memory/.

Pattern 4 — Seed Sharing: Agent extracts seed using seed-extraction skill → documents in 03_Memory/seeds/ → updates seeds/README.md.

VII. Token Efficiency

• Surgical reading: README (500-1000 tokens) + 2-3 relevant files (2000-5000 tokens each). Total ~5-10K tokens vs. 50K+ for reading everything.
• Incremental context: Start minimal (README + task), add files as needed, compress after milestones.
• Reference, don't copy: Link to source documents; extract 1-3 key sentences; never duplicate content.
• Metadata filtering: Check tags, status, and date before reading a file.

VIII. Maintenance

Weekly (30 min): Move finalized drafts, archive completed discussions, update decision log, refresh README priorities.

Monthly (1-2 hrs): Review specs for staleness, audit seeds for unused patterns, archive deprecated content, refactor structure as needed.

IX. Quality Checklist

• File is in the correct directory with proper naming convention
• Frontmatter (title, author, date, status, tags) is complete
• Purpose is clear and key insights are at the top
• Content is scannable (headings, lists, tables)
• Links to related content are included
• Workspace README updated if content is significant
• No unnecessary prose

Output

• File placed in the correct workspace directory with proper YYYY-MM-DD_topic-name_author.md naming
• Frontmatter block with title, author, date, status, tags, and related fields
• Related Content block linking to connected documents
• Updated workspace README.md when the new file is significant
• Grep results or directory listing when navigating to find existing content

Examples

Scenario 1: "Where does the new context compression proposal go?" → Placed at 00_Active/drafts/2026-04-08_context-compression-proposal_manus.md with complete frontmatter; README updated with a pointer.

Scenario 2: "Find the decision about which model router to use." → Grep across 00_Active/decisions/ and 01_Specifications/ for "model router"; returns file path and key excerpt from the decision's Rationale section.

Edge Cases

• Workspace has no README.md: create a minimal one before placing any new file; navigation depends on it
• Two agents writing to the same discussion file simultaneously: use threaded comment format with agent name + date prefix on each new block; never overwrite another agent's contribution
• A document spans multiple categories (e.g., a spec that also contains a decision): place in the primary category and link to it from the secondary directory's README

Anti-Patterns

• Placing files in the root of the workspace instead of the correct subdirectory — makes the repo unsearchable within three sessions
• Writing a document without frontmatter — breaks metadata filtering and forces full reads to determine relevance
• Copying full document text into another document instead of linking — doubles token cost every time both are loaded
• Creating thoughts.md or notes.md with no date or author — untraceable and unarchivable
• Skipping the README update after adding a significant file — the next agent starts blind

Related Skills

• handoff-protocol (agent-orchestration): Structures cross-session and cross-machine handoffs that land in the workspace. workspace-navigation defines where handoff documents go; handoff-protocol defines what they contain.
• decision-propagation (agent-orchestration): Produces the Decision Template referenced in Section V. Run after a workspace discussion reaches consensus to produce a properly formatted decision document for 00_Active/decisions/.
• context-ingestion (specification-driven-development): Reads workspace research and specs into working context for an implementation session — the natural downstream consumer of content placed by this skill.
• parallel-dispatch (agent-orchestration): When multiple agents write research or spec documents concurrently, workspace-navigation's naming convention and ownership rules prevent collisions across parallel tracks.