| name | fresh-start |
| description | Orient to project structure and load context. Use at the start of each new session or after context reset to understand the project state. |
| argument-hint | ["project-directory"] |
| allowed-tools | Read, Write, Edit, Glob, Grep, Bash, Skill, AskUserQuestion |
Orient to a scoped execution directory and load context for execution.
Workflow
Nested skill rule: When this workflow invokes another skill (e.g.,
/configure-verification, /phase-prep, /phase-start), you MUST return to
this checklist after the nested skill completes and continue with the next
unchecked item. Do not stop or summarize after a nested skill returns.
Copy this checklist and track progress:
Fresh Start Progress:
- [ ] Detect context (greenfield plan vs feature directory)
- [ ] Directory guard (verify AGENTS.md + EXECUTION_PLAN.md exist)
- [ ] Plan status guard (confirm this scoped directory is current)
- [ ] Git initialization (if needed)
- [ ] Feature branch setup (feature mode only)
- [ ] Scoped AGENTS check
- [ ] Read context and summarize
- [ ] Auto-configure verification (first run only)
- [ ] Phase state detection
- [ ] Auto-prep phase (first run only)
- [ ] Branch context detection
Project Directory
Use the current working directory by default.
If $1 is provided, treat $1 as the working directory and read files under $1 instead.
Context Detection
Determine working context before validation.
Convention: Run execution commands from the scoped directory that contains the active EXECUTION_PLAN.md:
- Greenfield:
plans/greenfield/
- Feature work:
features/<name>/
-
Let WORKING_DIR = $1 if provided, otherwise current working directory
-
If WORKING_DIR matches pattern */features/* (contains /features/ followed by a feature name):
- PROJECT_ROOT = parent of parent of WORKING_DIR (e.g.,
/project/features/foo → /project)
- SCOPE_DIR = WORKING_DIR
- MODE = "feature"
-
If WORKING_DIR matches pattern */plans/greenfield*:
- PROJECT_ROOT = parent of parent of WORKING_DIR (e.g.,
/project/plans/greenfield → /project)
- SCOPE_DIR = WORKING_DIR
- MODE = "greenfield"
-
Otherwise:
- PROJECT_ROOT = WORKING_DIR
- SCOPE_DIR = WORKING_DIR
- MODE = "greenfield-legacy"
Directory Guard (Wrong Directory Check)
Confirm the required files exist:
Plan Status Guard
Read PROJECT_ROOT/plans/PLAN_STATUS.md if it exists.
- If it exists, compare its
Current plan path to SCOPE_DIR relative to
PROJECT_ROOT.
- If
Current status is not active, STOP. Report that the current plan is
not implementable and ask the user to choose or reactivate a plan.
- If
Current plan points to a different path than SCOPE_DIR, STOP. Tell
the user the current active plan path and instruct them to cd there before
running /fresh-start.
- If
SCOPE_DIR is under plans/archive/ or features/archive/, STOP.
Archived plans are historical context only.
- If the manifest is missing, continue with legacy directory detection, but
report: "No plans/PLAN_STATUS.md found; proceeding from directory convention."
Git Initialization (First Run)
In PROJECT_ROOT (not the feature directory):
- Check whether this is already a git repo by running:
git -C PROJECT_ROOT rev-parse --is-inside-work-tree 2>/dev/null
If this returns "true", it's already a git repo.
- If not a git repo:
- If it is a git repo but has no commits yet:
Feature Branch Setup (Feature Mode Only)
If MODE = "feature", create an isolated branch for this feature work:
-
Derive FEATURE_NAME from the feature directory (basename of SCOPE_DIR, e.g., analytics-dashboard)
-
Check current branch:
git branch --show-current
-
If already on a feature/FEATURE_NAME branch, skip (already set up)
-
Otherwise, create and switch to the feature branch:
git add -A && git diff --cached --quiet || git commit -m "wip: uncommitted changes before feature/FEATURE_NAME"
Verify with git status that files are staged correctly.
git checkout -b feature/FEATURE_NAME
Verify with git branch that the new branch is active.
-
Report: "Created branch feature/FEATURE_NAME for isolated feature development"
Scoped AGENTS Check
If SCOPE_DIR/AGENTS.md exists and SCOPE_DIR != PROJECT_ROOT, report:
- "Scoped AGENTS.md found. Local instructions will layer on top of PROJECT_ROOT/AGENTS.md."
If it does not exist and MODE is feature or greenfield, report:
- "No scoped AGENTS.md found in this directory. Execution will fall back to PROJECT_ROOT/AGENTS.md only."
Auto-Configure Verification (First Run Only)
Silently auto-detect verification commands if not already configured.
- Read
PROJECT_ROOT/.claude/verification-config.json (use Read tool directly).
If the file does not exist (read fails with not found), treat as missing and go to step 4.
- If the file exists and has a
commands key → SKIP. Do not invoke /configure-verification.
Report "Verification config already exists" and go directly to Phase State Detection.
- If the file exists with
{"skipped": true} → SKIP. Same as above.
- If the file is missing, empty, or malformed (exists but has neither
commands nor
skipped), invoke /configure-verification with SCOPE_DIR. This runs silently
with no prompts. If it fails, report the error and continue.
→ CONTINUE to Phase State Detection (do not stop after this step).
Phase State Detection
Check for existing phase state to determine if this is a resume or first run:
-
Check if .claude/phase-state.json exists in:
SCOPE_DIR (if feature mode)
PROJECT_ROOT (if greenfield mode)
-
If valid phase state exists (file exists, parses correctly, has current_phase):
-
If no phase state exists (or file is invalid/stale):
- This is a first run. Continue to Auto-Prep Phase section.
Auto-Prep Phase (First Run Only)
Skip this section entirely if resuming (phase-state.json exists and is valid).
After context reading and verification config, automatically prepare the next phase:
-
Determine next phase number:
- If no phase state exists → Phase 1
- If phase state exists but is stale/invalid → Phase 1
-
Invoke /phase-prep {next_phase} silently.
- Phase-prep will verify prerequisites and auto-advance to
/phase-start if
all checks pass (via its existing auto-advance logic).
- If phase-prep blocks (human setup needed), it will report what's needed
and the user runs
/phase-start manually after resolving.
-
Fallback auto-advance check — After /phase-prep returns, read
.claude/phase-prep-result.json. If the file does NOT exist, phase-prep
dropped its auto-advance step (known issue with nested Skill tool invocations).
In this case:
→ CONTINUE to Branch Context Detection (do not stop after this step).
Branch Context Detection
Detect the current git branch and load relevant context:
-
Get current branch:
git branch --show-current 2>/dev/null
-
If branch matches feature/* pattern:
- Extract feature name from branch (e.g.,
feature/analytics-dashboard → analytics-dashboard)
- Look for matching feature directory:
PROJECT_ROOT/features/{feature-name}/
- If found and MODE is "greenfield", suggest: "Switch to feature mode? Found feature directory for this branch."
-
Summarize recent branch activity:
git log --oneline -5 2>/dev/null
Report: "Recent commits on this branch: {summary}"
-
Check for uncommitted changes:
git status --porcelain 2>/dev/null
If changes exist, report: "Note: {N} uncommitted changes in working tree"
Required Context
Read these files first:
- PROJECT_ROOT/AGENTS.md — Durable project-wide workflow guidelines
- PROJECT_ROOT/plans/PLAN_STATUS.md — Current plan manifest (if exists)
- SCOPE_DIR/AGENTS.md — Scoped execution guidance (if
SCOPE_DIR != PROJECT_ROOT)
- SCOPE_DIR/EXECUTION_PLAN.md — Tasks and acceptance criteria
Specification Documents
Check which of these exist and read them:
From PROJECT_ROOT (always check):
- LEARNINGS.md — Discovered patterns and gotchas (if exists)
From plans/greenfield/ (if greenfield mode):
- PRODUCT_SPEC.md — What we're building
- TECHNICAL_SPEC.md — How it's built
From PROJECT_ROOT (if MODE = "greenfield-legacy"):
- PRODUCT_SPEC.md — Legacy greenfield product spec
- TECHNICAL_SPEC.md — Legacy greenfield technical spec
From SCOPE_DIR (if feature mode):
- FEATURE_SPEC.md — Feature requirements
- FEATURE_TECHNICAL_SPEC.md — Feature technical approach
- FLOW_VERIFICATION_PLAN.md — Agent-runnable flow verification plan (if exists)
Your Task
- Read all available documents above
- Summarize your understanding:
- What is being built
- Current phase and progress
- Tech stack and key patterns
- Key learnings to follow (if LEARNINGS.md exists)
- Confirm you're ready to begin execution
Error Handling
| Situation | Action |
|---|
| Git init or clone failure | Report the error and stop — cannot proceed without a working repository. |
| Scoped AGENTS.md missing or malformed | Report the issue and continue with root AGENTS.md only. |
| /phase-prep failure | Report which pre-flight check failed and stop. Do not proceed to execution. |
Important: If LEARNINGS.md exists, apply those patterns throughout your work. These are project-specific conventions discovered during development that override general defaults.
Output Format (Programmatic — for calling skills)
When invoked by another skill (e.g., /go), return structured data:
{
"status": "<Status>",
"context_loaded": ["AGENTS.md", "EXECUTION_PLAN.md", "FLOW_VERIFICATION_PLAN.md"],
"blockers": [],
"next_action": "Continue with Phase 1, Task 1"
}
Status enum — exactly one of:
| Value | Meaning |
|---|
ready | All context loaded, no blockers. Execution can proceed. |
blocked | One or more blockers prevent execution (missing files, failed git init, etc.). |
partial | Context partially loaded; some optional files missing but execution can proceed with caveats. |
Field descriptions:
| Field | Type | Description |
|---|
status | Status | Overall readiness result (see enum above). |
context_loaded | string[] | List of context files successfully read (e.g., "AGENTS.md", "EXECUTION_PLAN.md"). |
blockers | string[] | Issues preventing execution. Empty array when status is ready or partial. |
next_action | string | Suggested next step (e.g., "Continue with Phase 2, Task 3" or "Run /setup first"). |