manifest-implement - SKILL.md Agent Skill

name: manifest-implement description: | Structured implementation flow for Manifest features — from investigation through execution with subagents. Follows the full chain: investigate → clarify → explore → validate → plan → todos → branch → execute → review. No shortcuts. Invoke manually when building a new feature or significant change in a Manifest project.

Manifest Implement

A structured implementation flow for building features in a Manifest project. This is the opinionated way to go from idea to shipped code — investigation, design, planning, execution, and review.

Announce at start: "Starting a Manifest implementation session. Let me investigate the codebase first, then we'll work through this step by step."

⚠️ MANDATORY: No Skipping Without Permission

You MUST follow all phases. Your judgment that something is "simple" or "straightforward" is NOT sufficient to skip steps.

The ONLY exception: The user explicitly says something like:

"Skip the plan, just implement it"
"Just do it quickly"
"No need for the full process"

If the user hasn't said this, you follow the full flow. Period.

Why this matters: You will be tempted to rationalize. You'll think "this is just a small feature" or "this is obvious, no plan needed." That's exactly when the process matters most — consistency builds trust, and "small" changes have a way of growing.

The Flow

Phase 1: Investigate the Manifest Project
    ↓
Phase 2: Clarify Requirements
    ↓
Phase 3: Explore Approaches (Manifest-aware)
    ↓
Phase 4: Present & Validate Design
    ↓
Phase 5: Write Plan
    ↓
Phase 6: Create Todos
    ↓
Phase 6.5: Create Feature Branch
    ↓
Phase 7: Execute with Subagents (scout → workers → review)
    ↓
Phase 8: Review

🛑 STOP — Before Writing Any Code

If you're about to edit or create source files, STOP and check:

✅ Did you complete Phase 4 (design validation)?
✅ Did you write a plan to ~/.pi/history/<project>/plans/?
✅ Did you create todos?

If any answer is NO and the user didn't explicitly skip → you're cutting corners. Go back.

Phase 1: Investigate the Manifest Project

Before asking questions, understand the codebase. This is a Manifest project — you know the structure.

Read the essentials

# Understand the app
cat VISION.md
cat AGENTS.md

# See what exists
ls features/
ls schemas/
ls services/
ls extensions/
ls config/

# Check recent activity
git log --oneline -n 10

Look at existing patterns

# Read 2-3 existing features to understand the app's patterns
head -40 features/*.ts | head -120

# Check schemas for the data model
ls schemas/
head -30 schemas/*.ts | head -90

# Check what services exist
ls services/

What you're looking for

What does the app do? — VISION.md tells you the purpose
What features exist? — Understand the current surface area
What's the data model? — Schemas tell you what entities exist
What services are shared? — Don't reinvent existing logic
What conventions does this specific project use? — Every Manifest project follows the framework but may have its own patterns on top

After investigating, share what you found:

"Here's what I see: [app purpose from VISION.md]. The app currently has [N] features covering [areas]. The data model includes [key tables]. Existing services handle [what]. Now let me understand what you're looking to build."

Phase 2: Clarify Requirements

Work through requirements one topic at a time:

Topics to Cover

Purpose — What problem does this solve? Who's it for?
Scope — What's in? What's explicitly out?
Manifest mapping — Is this a feature, a schema change, a service, or a combination?
Constraints — Performance, compatibility, timeline?
Success criteria — How do we know it's done?

How to Ask

Group related questions, use /answer for multiple questions
Prefer multiple choice when possible (easier to answer)
Don't overwhelm — if you have many questions, batch them logically

[After listing your questions]
execute_command(command="/answer", reason="Opening Q&A for requirements")

Manifest-Specific Questions to Consider

Does this need a new schema, or does it extend an existing one?
Should this be a single feature or multiple features (one behavior per file)?
Is there authentication required? What level?
What side effects does this have? (Database writes, external API calls, emails)
What are the error cases? (Think in HTTP status codes)
Does this share logic with existing features? (Service extraction)
Does this need rate limiting?

Keep Going Until Clear

After each round of answers, either:

Ask follow-up questions if something is still unclear
Summarize your understanding and confirm: "So we're building X that does Y for Z. Right?"

Don't move to Phase 3 until requirements are clear.

Phase 3: Explore Approaches (Manifest-Aware)

Once requirements are understood, propose 2-3 approaches. Every approach should be grounded in Manifest conventions.

Think in Manifest Primitives

Every feature maps to one or more of these:

Feature (defineFeature()) — The behavior. One file per behavior.
Schema (pgTable()) — The data. One file per table.
Service (plain object) — Shared logic. Only when two features need it.
Extension — If it's reusable across projects.

Present Options

"A few ways we could approach this:

Simple approach — [description].

Features: [list feature files]

Schemas: [new/modified tables]

Pros: fast, minimal

Cons: less flexible

Full approach — [description].

Features: [list feature files]

Schemas: [new/modified tables]

Services: [shared logic]

Pros: extensible

Cons: more work

I'd lean toward #1 because [reason]. What do you think?"

Key Principles

Lead with your recommendation — don't make them guess
Be explicit about tradeoffs — every choice has costs
YAGNI ruthlessly — remove unnecessary complexity
One feature, one file — if you're putting two behaviors in one file, rethink
Don't extract services prematurely — wait for a second consumer
Trust the framework — don't reach for external libraries when manifest/ handles it

Phase 4: Present & Validate Design

Present the design in sections, validating each before moving on.

Why Sectioned?

A wall of text gets skimmed; sections get read
Catches misalignment early
Easier to course-correct than rewrite

Section by Section

Keep each section to 200-300 words.

Section 1: Feature Breakdown

Which features we're creating, what each does, their routes:

"Does this breakdown make sense? Any behaviors missing or combined that shouldn't be?"

Section 2: Data Model

New or modified schemas, relationships:

"Does this data model capture everything we need?"

Section 3: Logic Flow

How requests flow through features, what services are involved:

"Does this flow make sense?"

Section 4: Error Handling & Edge Cases

What can go wrong, how we handle it, HTTP status codes:

"Any edge cases I'm missing?"

Section 5: Testing Approach

How we'll verify it works:

"Does this testing approach give you confidence?"

Not every project needs all sections — use judgment based on complexity. A simple single-feature change might only need Sections 1 and 3.

Incorporating Feedback

If they suggest changes:

Acknowledge the feedback
Update your understanding
Re-present that section if needed
Continue to next section

Phase 5: Write Plan

Once the design is validated:

"Design is solid. Let me write up the plan."

Create: ~/.pi/history/<project>/plans/YYYY-MM-DD-[plan-name].md

<project> = basename of the current working directory (e.g., my-app for /Users/haza/Projects/my-app)

Write the Full Plan

By this point, the design has been explored, validated, and refined through Phases 2–4. Don't re-ask for approval on every section — just write the complete plan and present it.

# [Plan Name]

**Date:** YYYY-MM-DD
**Status:** Draft
**Directory:** /path/to/project

## Overview

[What we're building and why — 2-3 sentences]

## Goals

- Goal 1
- Goal 2
- Goal 3

## Approach

[High-level technical approach]

### Features

| Feature | Route | Description |
|---------|-------|-------------|
| feature-name | POST /api/path | What it does |

### Schemas

| Table | Purpose |
|-------|---------|
| table_name | What it stores |

### Services

| Service | Purpose |
|---------|---------|
| serviceName | What it provides (only if needed) |

### Key Decisions

- Decision 1: [choice] — because [reason]
- Decision 2: [choice] — because [reason]

## Dependencies

- Libraries needed (if any beyond what Manifest provides)

## Risks & Open Questions

- Risk 1
- Open question 1

After writing, briefly confirm:

"Plan is written. Ready to create the todos, or anything you want to adjust?"

Phase 6: Create Todos

After the plan is verified, break it into todos.

Make Todos Bite-Sized

Each todo = one focused action (2-5 minutes).

❌ Too big: "Implement user registration"

✅ Granular:

"Create schemas/users.ts with users table"
"Create features/register-user.ts with validation and insert"
"Create tests/register-user.test.ts with happy path and error cases"
"Commit: 'feat(users): add user registration with email validation'"

Manifest-Specific Todo Templates

Feature todo:

Plan: ~/.pi/history/<project>/plans/YYYY-MM-DD-plan-name.md

## Task
Create feature `feature-name` at `features/feature-name.ts`

## Details
- Route: POST /api/path
- Authentication: required
- Input: field1 (string, required), field2 (integer, optional)
- Side effects: Inserts row into table_name
- Error cases: 409 if duplicate, 422 if validation fails

## Acceptance Criteria
- [ ] Uses defineFeature() with description, sideEffects, errorCases
- [ ] Every input field has a description
- [ ] handle() is linear — no deeply nested logic
- [ ] Returns standard envelope responses via ok()/fail()

Schema todo:

Plan: ~/.pi/history/<project>/plans/YYYY-MM-DD-plan-name.md

## Task
Create schema `schemas/tableName.ts`

## Details
- Table: table_name
- Columns: id (uuid, pk), name (varchar), created_at (timestamp)
- Relationships: references other_table(id)

## Acceptance Criteria
- [ ] JSDoc on the table explaining what it stores
- [ ] JSDoc on every column explaining what it's for
- [ ] Uses Drizzle ORM patterns consistently

Service todo:

Plan: ~/.pi/history/<project>/plans/YYYY-MM-DD-plan-name.md

## Task
Create service `services/serviceName.ts`

## Details
- Methods: doThing(input) → Result
- Used by: feature-a, feature-b

## Acceptance Criteria
- [ ] Plain exported object with methods — no class
- [ ] JSDoc on every method
- [ ] Typed parameters — no any

Creating Todos

todo(action: "create", title: "Task 1: [description]", tags: ["plan-name"], body: "...")

Phase 6.5: Create Feature Branch

Always create a feature branch before executing. Never work directly on main.

git checkout -b feat/<short-descriptive-name>

Branch naming:

feat/<name> for features
fix/<name> for bug fixes
refactor/<name> for refactors

Keep the name short and descriptive (e.g., feat/user-registration, fix/null-email, refactor/auth-service).

Phase 7: Execute with Subagents

Always start with a scout, then run workers sequentially. The scout gathers context about all relevant files upfront so workers spend less time on discovery.

The Pattern

Run scout first — gathers context about all files relevant to the plan
Run worker for each todo — one at a time, each referencing the scout's context
Check results — verify files were created/modified correctly
Handle failures — if a worker fails, diagnose and retry or fix manually
Run reviewer last — only after all todos are complete

Why Scout First?

Workers are expensive (Sonnet). Every minute a worker spends grepping and reading files to orient itself is wasted. The scout (Haiku) is fast and cheap — it maps out the codebase, identifies key files, notes patterns and conventions, and hands that context to each worker.

Example

// Step 1: Scout gathers context for the entire plan
{ agent: "scout", task: "Gather context for implementing [feature] in this Manifest project. Read the plan at ~/.pi/history/<project>/plans/YYYY-MM-DD-feature.md. Read VISION.md and AGENTS.md. Identify all files that will be created or modified. Map out existing features, schemas, services patterns that workers will need to follow." }

// Step 2: Workers execute todos sequentially
{ agent: "worker", task: "Implement TODO-xxxx. Use the commit skill to write a polished, descriptive commit message. Mark the todo as done. Plan: ~/.pi/history/<project>/plans/YYYY-MM-DD-feature.md\n\nScout context:\n{paste scout context}" }

// More workers for each todo...

// After all todos, review
{ agent: "reviewer", task: "Review the feature branch against main. Plan: ~/.pi/history/<project>/plans/YYYY-MM-DD-feature.md" }

Practical Implementation

The scout writes its findings to ~/.pi/history/<project>/context.md. Before spawning each worker, read the scout's context file and paste it into the worker's task:

// 1. Run scout
subagent({ agent: "scout", task: "Gather context for [feature]. Read the plan at ~/.pi/history/<project>/plans/YYYY-MM-DD-feature.md. Read VISION.md and AGENTS.md for project context. Map existing features, schemas, services, and conventions." })

// 2. Read the scout's output
const scoutContext = read("~/.pi/history/<project>/context.md")

// 3. Pass it to each worker
subagent({ agent: "worker", task: `Implement TODO-xxxx. Use the commit skill to write a polished, descriptive commit message. Mark the todo as done. Plan: ~/.pi/history/<project>/plans/YYYY-MM-DD-feature.md

Scout context (your starting baseline — still do targeted discovery as needed):
${scoutContext}` })

What the Scout Should Cover

Tell the scout to focus on:

VISION.md and AGENTS.md — app purpose and project conventions
Files from the plan — read the plan, identify every file mentioned
Existing features — how similar features are structured in this project
Schemas — current data model, relationships
Services — what shared logic already exists
manifest/ framework — relevant framework source if the plan touches it

Handling Reviewer Findings

When the reviewer returns with issues, act on the important ones:

Triage the findings:
- P0 (Drop everything) — Real bugs, security holes, data loss. Must fix now.
- P1 (Foot gun) — Genuine traps, real maintenance dangers. Should fix.
- P2 (Worth mentioning) — Fix if quick, otherwise note for later.
- P3 (Almost irrelevant) — Skip.

Create todos for P0s and P1s:

todo({ action: "create", title: "Fix: [issue from reviewer]", body: "..." })

Kick off workers to fix them:

{ agent: "worker", task: "Fix TODO-xxxx (from review). Use the commit skill. Mark todo as done." }

Don't re-review minor fixes — only run reviewer again if fixes were substantial.

⚠️ MANDATORY: Always Run Reviewer

After all workers complete, you MUST run the reviewer. No exceptions. The workflow is not complete until the reviewer has run.

// This is NOT optional. Always end with:
{ agent: "reviewer", task: "Review the feature branch against main. Plan: ~/.pi/history/<project>/plans/YYYY-MM-DD-feature.md" }

⚠️ Avoid Parallel Workers in Git Repos

Do NOT use parallel workers when they share a git repository. Workers that commit to the same repo will conflict. Always run workers sequentially.

When parallel IS safe:

Read-only tasks (e.g., multiple scouts gathering info)
Workers operate on completely separate git repos

Alternative: Same Session

If the user prefers hands-on work:

"Would you rather I work through these myself while you review?"

Then work through todos sequentially:

Claim the todo
Implement
Verify
Commit using the manifest-commit skill (polished, descriptive message)
Close the todo
Move to next

Phase 8: Review (Mandatory)

The reviewer agent is Manifest-aware. It checks:

Implementation matches the plan
Manifest conventions are followed (descriptions, sideEffects, errorCases, JSDoc)
Code quality and correctness
Security considerations

🛑 STOP — Before Reporting Completion

Check:

✅ All worker todos are closed?
✅ Every completed todo has a polished commit (using the manifest-commit skill)?
✅ Reviewer has run? ← If no, run it now
✅ Reviewer findings triaged and addressed?

Do NOT tell the user the work is done until all four are true.

Do NOT squash merge or merge the feature branch into main. The feature branch stays as-is with its individual, well-crafted commits.

Commit Strategy

Do NOT squash merge or merge feature branches back into main. Every completed todo gets its own polished, descriptive commit on the feature branch using the manifest-commit skill — always, no exceptions.

What Makes a Good Manifest Commit

Load the manifest-commit skill every time. Commits in Manifest are knowledge transfer — the primary way agents and humans understand what happened. A reader of git log should understand the change without looking at the diff.

Subject line: <type>(<scope>): <summary>, <= 72 chars
Body: What changed, why, how, which files, any migration notes
Be thorough — not just "implement X" but explain the approach and rationale

Working with Todos During Implementation

Claiming

todo(action: "claim", id: "TODO-xxxx")

Claim when you start working. Don't claim if sub-agents will pick it up.

Progress Notes

todo(action: "append", id: "TODO-xxxx", body: "Implemented the validation logic...")

Closing

todo(action: "update", id: "TODO-xxxx", status: "closed")

Viewing

/todos — visual todo manager
todo(action: "list") — open and assigned
todo(action: "get", id: "TODO-xxxx") — full details

Tips

Read the Room

If they have a clear vision → validate rather than over-question
If they're eager to start → move faster through phases (but still hit all phases)
If they're uncertain → spend more time exploring

Stay Conversational

This is a dialogue, not an interrogation
Phases can be quick depending on complexity, but don't skip them

Be Opinionated

You know Manifest. Share your perspective on the right way to do things.
"I'd suggest a single feature for this because..." is more helpful than listing options
Push back if something violates Manifest conventions

Keep It Focused

One topic at a time
Don't let scope creep during planning
Parking lot items for later: "Good thought — let's note that for v2"

Trust the Framework

Don't add dependencies when manifest/ already handles it
Don't create abstractions when a linear handle() works
Don't extract services until you have two consumers
Read the framework source if you're unsure what it provides