Start

• Core principles (Read Before Write, Follow Standards, etc.)
• File system structure
• Development process
• Best practices

Step 2: Get Current Context

python3 ./.trellis/scripts/get_context.py

This shows: developer identity, git status, current task (if any), active tasks.

Step 3: Read Guidelines Index

python3 ./.trellis/scripts/get_context.py --mode packages

This shows available packages and their spec layers. Read the relevant spec indexes:

cat .trellis/spec/<package>/<layer>/index.md   # Package-specific guidelines
cat .trellis/spec/guides/index.md              # Thinking guides (always read)

Important: The index files are navigation — they list the actual guideline files (e.g., error-handling.md, conventions.md, mock-strategies.md). At this step, just read the indexes to understand what's available. When you start actual development, you MUST go back and read the specific guideline files relevant to your task, as listed in the index's Pre-Development Checklist.

Step 4: Report and Ask

Report what you learned and ask: "What would you like to work on?"

Task Classification

When user describes a task, classify it:

Classification Signals

Trivial/Simple indicators:

• User specifies exact file and change
• "Fix the typo in X"
• "Add field Y to component Z"
• Clear acceptance criteria already stated

Complex indicators:

• "I want to add a feature for..."
• "Can you help me improve..."
• Mentions multiple areas or systems
• No clear implementation path
• User seems unsure about approach

Decision Rule

If in doubt, use Brainstorm + Task Workflow.

Task Workflow ensures specs are injected to agents, resulting in higher quality code. The overhead is minimal, but the benefit is significant.

Question / Trivial Fix

For questions or trivial fixes, work directly:

1. Answer question or make the fix
2. If code was changed, remind user to run $finish-work

Simple Task

For simple, well-defined tasks:

1. Quick confirm: "I understand you want to [goal]. Shall I proceed?"
2. If no, clarify and confirm again
3. If yes: execute ALL steps below without stopping. Do NOT ask for additional confirmation between steps.
   • Create task directory (Phase 1 Path B, Step 2)
   • Write PRD (Step 3)
   • Research codebase (Phase 2, Step 5)
   • Configure context (Step 6)
   • Activate task (Step 7)
   • Implement (Phase 3, Step 8)
   • Check quality (Step 9)
   • Complete (Step 10)

Complex Task - Brainstorm First

For complex or vague tasks, automatically start the brainstorm process — do NOT skip directly to implementation.

See the $brainstorm skill for the full process. Summary:

1. Acknowledge and classify - State your understanding
2. Create task directory - Track evolving requirements in prd.md
3. Ask questions one at a time - Update PRD after each answer
4. Propose approaches - For architectural decisions
5. Confirm final requirements - Get explicit approval
6. Proceed to Task Workflow - With clear requirements in PRD

Subtask Decomposition: If brainstorm reveals multiple independent work items, consider creating subtasks using --parent flag or add-subtask command. See the brainstorm skill's Step 8 for details.

Task Workflow (Development Tasks)

Why this workflow?

• Research Agent analyzes what specs are needed
• Specs are configured in jsonl files
• Implement Agent receives specs via Hook injection
• Check Agent verifies against specs
• Result: Code that follows project conventions automatically

Overview: Two Entry Points

From Brainstorm (Complex Task):
  PRD confirmed → Research → Configure Context → Activate → Implement → Check → Complete

From Simple Task:
  Confirm → Create Task → Write PRD → Research → Configure Context → Activate → Implement → Check → Complete

Key principle: Research happens AFTER requirements are clear (PRD exists).

Phase 1: Establish Requirements

Path A: From Brainstorm (skip to Phase 2)

PRD and task directory already exist from brainstorm. Skip directly to Phase 2.

Path B: From Simple Task

Step 1: Confirm Understanding [AI]

Quick confirm:

• What is the goal?
• What type of development? (frontend / backend / fullstack)
• Any specific requirements or constraints?

Step 2: Create Task Directory [AI]

TASK_DIR=$(python3 ./.trellis/scripts/task.py create "<title>" --slug <name>)

Step 3: Write PRD [AI]

Create prd.md in the task directory with:

# <Task Title>

## Goal
<What we're trying to achieve>

## Requirements
- <Requirement 1>
- <Requirement 2>

## Acceptance Criteria
- [ ] <Criterion 1>
- [ ] <Criterion 2>

## Technical Notes
<Any technical decisions or constraints>

Phase 2: Prepare for Implementation (shared)

Both paths converge here. PRD and task directory must exist before proceeding.

Step 4: Code-Spec Depth Check [AI]

If the task touches infra or cross-layer contracts, do not start implementation until code-spec depth is defined.

Trigger this requirement when the change includes any of:

• New or changed command/API signatures
• Database schema or migration changes
• Infra integrations (storage, queue, cache, secrets, env contracts)
• Cross-layer payload transformations

Must-have before proceeding:

• Target spec files to update are identified
• Concrete contract is defined (signature, fields, env keys)
• Validation and error matrix is defined
• At least one Good/Base/Bad case is defined

Step 5: Research the Codebase [AI]

Based on the confirmed PRD, call Research Agent to find relevant specs and patterns:

Task(
  subagent_type: "research",
  prompt: "Analyze the codebase for this task:

  Task: <goal from PRD>
  Type: <frontend/backend/fullstack>

  Please find:
  1. Relevant spec files in .trellis/spec/
  2. Existing code patterns to follow (find 2-3 examples)
  3. Files that will likely need modification

  Output:
  ## Relevant Specs
  - <path>: <why it's relevant>

  ## Code Patterns Found
  - <pattern>: <example file path>

  ## Files to Modify
  - <path>: <what change>",
  model: "opus"
)

Step 6: Configure Context [AI]

Initialize default context:

python3 ./.trellis/scripts/task.py init-context "$TASK_DIR" <type>
# type: backend | frontend | fullstack

Add specs found by Research Agent:

# For each relevant spec and code pattern:
python3 ./.trellis/scripts/task.py add-context "$TASK_DIR" implement "<path>" "<reason>"
python3 ./.trellis/scripts/task.py add-context "$TASK_DIR" check "<path>" "<reason>"

Step 7: Activate Task [AI]

python3 ./.trellis/scripts/task.py start "$TASK_DIR"

This sets .current-task so hooks can inject context.

Phase 3: Execute (shared)

Step 8: Implement [AI]

Call Implement Agent (specs are auto-injected by hook):

Task(
  subagent_type: "implement",
  prompt: "Implement the task described in prd.md.

  Follow all specs that have been injected into your context.
  Run lint and typecheck before finishing.",
  model: "opus"
)

Step 9: Check Quality [AI]

Call Check Agent (specs are auto-injected by hook):

Task(
  subagent_type: "check",
  prompt: "Review all code changes against the specs.

  Fix any issues you find directly.
  Ensure lint and typecheck pass.",
  model: "opus"
)

Step 10: Complete [AI]

1. Verify lint and typecheck pass
2. Report what was implemented
3. Remind user to:
   • Test the changes
   • Commit when ready
   • Run $record-session to record this session

Continuing Existing Task

If get_context.py shows a current task:

1. Read the task's prd.md to understand the goal
2. Check task.json for current status and phase
3. Ask user: "Continue working on <task-name>?"

If yes, resume from the appropriate step (usually Step 7 or 8).

Skills Reference

User Skills [USER]

AI Scripts [AI]

Sub Agents [AI]

Key Principle

Specs are injected, not remembered.

The Task Workflow ensures agents receive relevant specs automatically. This is more reliable than hoping the AI "remembers" conventions.