markshust/hcf/plan-create

Plan Create

Transform feature requests into structured plans with task breakdowns, dependencies, and TDD requirements for parallel autonomous execution.

Project Architecture

<architecture> @.claude/architecture.md </architecture>

Execution Flow

Phase 0: Pre-plan hook

Run this once, at the very start, before Phase 1 discovery.

Run the `pre-plan` hook.

Run the discovery script — never enumerate agent files by hand:

"{skill-base-dir}/../../hooks/discover-hooks.sh" --hook=pre-plan

{skill-base-dir} is the `Base directory for this skill` value stated when this skill loaded. ${CLAUDE_PLUGIN_ROOT} is not available in a skill's Bash calls.

• Exit 0, empty stdout → empty hook. This is the default. Return immediately: print nothing, spawn nothing, say nothing about the hook at all, and proceed to Phase 1. Do not narrate that the script returned no agents — that is the narration HOOKS.md forbids.
• Exit 0, output → print the output verbatim as the resolved agent order, then spawn each listed agent in order per its mode.
• Any non-zero exit → stop. Surface the script's stderr verbatim and do not continue planning. Exit 3 means an agent file declares an invalid phase or mode; exit 4 means enrollment drifted. Neither is an empty hook. If the script is missing or not executable, that is also a hard stop — there is no fallback, and you must not reconstruct the routine by hand.

Capture the enrollment fingerprint now, for Phase 6's drift check:

"{skill-base-dir}/../../hooks/discover-hooks.sh" --fingerprint

Hold that value. A fingerprint is only ever produced by running the script — never reconstruct, abbreviate, or recall one from memory. If it is not to hand at Phase 6, re-run the command and say so. A hallucinated digest makes every later check fail with bogus drift and bricks the plan.

IMPORTANT — no plan name exists yet. The pre-plan hook fires before Phase 1 discovery. The plan name and plan directory are not created until Phase 3, so there is no plan name to pass. Each pre-plan agent receives only:

1. The raw feature request (the user's verbatim ask).
2. The project's architecture context:

## Project Architecture
{paste the COMPLETE content of <architecture> verbatim}

Do not attempt to pass a plan name or plan directory to a pre-plan agent — neither exists yet. The pre-plan hook is thin by design (policy gate / house-context injection).

Phase 1: Discovery & Assumption Brainstorm

Approach this phase like a solution architect meeting with a client to flush out scope and requirements. Your job is not to ask everything — it's to think hard about the shape of the solution before asking grounded questions. Find the integration points, then enumerate the design axes the user probably hasn't thought about.

Quick scope check first. If the ask already specifies data model, scope boundaries, and integration points (e.g., "Add an email_verified_at column to users and a SendVerificationEmail job"), do a brief existence-check (do the named files/models exist?) and skip ahead to Phase 2 — the brainstorm below is for under-specified asks.

1. Codebase discovery

Read files specifically related to the ask. Architecture context is already loaded via the <architecture> block at the top of this skill — do not re-read .claude/architecture.md. Instead:

• Glob for domain-related files (e.g., for "track books", look for book*, *Book*, library/reading-list files)
• Read existing models, controllers, or components that overlap with the ask
• Note existing patterns the new feature must conform to or extend
• Identify what's already built vs. greenfield

2. Permutation & assumption brainstorm

For each noun and verb in the ask, enumerate plausible interpretations. Then enumerate the hidden axes — design decisions the user almost certainly hasn't specified but the implementer needs to know. Surface the non-obvious ones a senior engineer would catch and a junior would miss.

Focus on permutations that meaningfully change scope, data model, or architecture. Avoid minutiae (button colors, naming bikesheds, trivial config defaults).

Example for "track books I'm reading":

• Book schema: title only? +author/ISBN/cover/genre/pages/publication date?
• List structure: one global list, or multiple states ("want to read", "reading", "finished")?
• "Finished" semantics: boolean state, dated event, or reversible?
• Hidden axes the ask didn't mention: auth/multi-user, persistence layer, edit/delete operations, search/filter/sort, ratings/notes, web vs mobile UI, external lookups (ISBN APIs).

3. Diff against codebase

Produce a short "what I found vs. what you asked" comparison. Examples:

• "You have a User model with auth — books would extend it for per-user lists"
• "No models exist yet — this is greenfield"
• "Your existing controllers follow a ResourceController pattern; book routes would conform"

Issue Detection: If the user references a GitHub issue (e.g., "#18", "issue 18", a GitHub issue URL), capture it for the ## Related Issues field in _plan.md. Use Closes #N for issues that will be fully resolved by this plan, or Relates to #N for partial/tangential references. If no issue is mentioned, set the field to "none".

Phase 2: Grounded Clarification

Surface findings to the user, then ask categorized questions. The goal is to flush out enough to write a confident plan — like a solution architect leaving a client meeting with the spec they need.

Format your response so users can skim past findings if they want — the Questions section must be self-sufficient.

Present in this order:

What I Found

- {codebase findings: existing models, patterns, integration points, greenfield vs. extension}

Key Permutations to Resolve

{1-3 sentences naming the design axes that meaningfully shape scope/data model/architecture, derived from the brainstorm}

Questions

Must answer (these shape scope, data model, or architecture):

1. {scope-shaping question}

2. {data-model question}

3. {integration question}

Will default if you don't specify (defaults stated):

- {dimension}: {proposed default}

- {dimension}: {proposed default}

Question quality guidance:

• No hard cap on count — some plans need many questions, some need few. Ask as many as warranted.
• Stay focused on scope, data model, integration, and architecture. Skip minutiae (UI colors, exact field names, naming bikesheds, trivial config).
• Each "must answer" should be a decision that, if assumed wrong, would force a rewrite. Each "will default" should be safe to assume with a stated default the user can override.
• If the Phase 1 quick scope check determined the ask was already concrete, skip to Phase 3 with no questions asked.

Phase 3: Define the Plan

Once you understand the requirements, create the plan overview:

Create feature branch:

git checkout -b feature/{plan-name}

If the branch already exists (resuming a plan), check it out instead:

git checkout feature/{plan-name}

Resolve the plans directory, once, and reuse $PLANS_DIR for every path below. It is .claude/plans unless the project overrides it in .claude/hcf.json (see Plans Directory in the README). The script answers with an absolute path resolved from the project root, so the result does not depend on the working directory:

PLANS_DIR="$("{skill-base-dir}/../../hooks/resolve-plans-dir.sh")" || exit 1

Never substitute a hand-written `jq` line or a literal `.claude/plans`. A non-zero exit means the project root or the configured value is unusable; surface the script's stderr verbatim and stop rather than falling back to the default — writing plans somewhere the project did not ask for is the failure this indirection exists to prevent.

Create plan directory:

mkdir -p "$PLANS_DIR/{plan-name}"

Use a kebab-case name derived from the feature (e.g., user-authentication, payment-processing). Quote every path built from $PLANS_DIR — a configured directory may contain spaces.

Record the enrollment fingerprint so plan-orchestrate can detect that hook enrollment changed between planning and execution. Write it by redirecting a fresh invocation — do not re-type the value captured at Phase 0:

"{skill-base-dir}/../../hooks/discover-hooks.sh" --fingerprint > "$PLANS_DIR/{plan-name}/.hook-fingerprint"

The file holds exactly that one line and nothing else — no heading, no comment, no explanation. Writing it as a redirect rather than transcribing a remembered value is deliberate: Phases 1–2 are a long interactive stretch, and a value carried across them is exactly what gets lost and then invented.

Create `_plan.md`:

# Plan: {Feature Name}
 
## Created
{date}
 
## Status
planning | ready | in_progress | completed | blocked
 
## Objective
{1-2 sentence description of what this plan achieves}
 
## Related Issues
{list of GitHub issue references, e.g., "Closes #18", "Relates to #42", or "none"}
 
## Discovery Notes
{Brief summary of Phase 1 findings: existing models/patterns to extend, integration points, greenfield vs. existing code, key assumptions resolved during clarification. Captures context for future readers and resumed sessions.}
 
## Scope
 
### In Scope
-{bullet points of included functionality}
 
### Out of Scope
-{bullet points of explicitly excluded functionality}
 
## Success Criteria
-[ ] {measurable outcome 1}
-[ ] {measurable outcome 2}
-[ ] All tests passing
-[ ] Code follows project standards
 
## Task Overview
| Task | Description | Depends On | Status |
|------|-------------|------------|--------|
| 001 | {title} | - | pending |
| 002 | {title} | 001 | pending |
| ... | ... | ... | ... |
 
## Architecture Notes
{Any architectural decisions or patterns to follow}
 
## Risks & Mitigations
-{potential risk}: {mitigation strategy}

Phase 4: Break Down into Tasks

Create numbered task files. Follow these principles:

Task Sizing:

• Each task should be completable in 1 TDD cycle (15-30 min of focused work)
• A task typically has 3-7 requirements
• If a task has >7 requirements, split it
• If a task has <3 requirements, consider combining with related task

Dependency Rules:

• Tasks with no dependencies can run in parallel
• Explicitly declare ALL dependencies (not just immediate ones)
• Avoid circular dependencies
• Maximize parallelism by minimizing unnecessary dependencies

Required Tasks:

• If the plan creates a new package, the final task must create a README.md following the project's Package README Standards (see code-standards.md). This task depends on all other tasks so the README accurately reflects what was built.

Task File Format (`{NNN}-{task-name}.md`):

# Task {NNN}: {Title}
 
**Status**: pending
**Depends on**: [{comma-separated task numbers, or "none"}]
**Retry count**: 0
 
## Description
{2-4 sentences describing what this task accomplishes and why}
 
## Context
{Any relevant context the implementer needs to know}
-Related files: {list key files to modify or reference}
-Patterns to follow: {reference to existing patterns in codebase}
 
## Requirements (Test Descriptions)
Write requirements as exact test names. These become the test method names.
 
-[ ] `it creates a new user with valid email and password`
-[ ] `it rejects duplicate email addresses with validation error`
-[ ] `it hashes passwords before storing in database`
-[ ] `it returns user object with id after successful creation`
 
## Acceptance Criteria
-All requirements have passing tests
-Code follows code standards
-No decrease in test coverage
 
## Implementation Notes
(Left blank - filled in by programmer during implementation)

Phase 5: Validate Dependencies