Skip to content
Agent Orchestration
Command

/custom-plan

Create a structured feature plan using PLAN + WORKING_MEMORY + USER_VERIFICATION (tasklist) files

From plugin
dexto
65214 skills7 agents14 commands

How it fires

How this command gets triggered: by you, by Claude, or both.

  • Fires itselfClaude auto-loads it when your prompt matches the work.
  • You can call itInvoke it directly when you want it.
  • Slash command/custom-plan

Context preview

What this command does when you run it.

Create a structured feature plan using PLAN + WORKING_MEMORY + USER_VERIFICATION (tasklist) files

Command definition

custom-plan.md
description: "Create a structured feature plan using PLAN + WORKING_MEMORY + USER_VERIFICATION (tasklist) files"
allowed-tools: ["bash", "read", "glob", "grep", "task"]

Custom Plan (PLAN + Working Memory + Tasklist)

Create feature plans using the PLAN + Working Memory + Tasklist paradigm:

  • `PLAN.md` is the source of truth (problem, goals, design, and a concrete tasklist).
  • `WORKING_MEMORY.md` is the live scratchpad agents update while executing.
  • `USER_VERIFICATION.md` is the owner-only decision + manual verification checklist.

1) Create the plan directory + files

Create a new directory:

  • `feature-plans/<topic>/`

Add these three files:

  • `feature-plans/<topic>/PLAN.md`
  • `feature-plans/<topic>/WORKING_MEMORY.md`
  • `feature-plans/<topic>/USER_VERIFICATION.md`

2) Write `PLAN.md` (durable, structured)

At the top, link the other two files and state the update rules:

# <Title>

**Working memory:** [`WORKING_MEMORY.md`](./WORKING_MEMORY.md) — read before starting work; update after each task.
**Owner verification list:** [`USER_VERIFICATION.md`](./USER_VERIFICATION.md) — add entries for owner-only decisions or manual checks; clear before shipping.

Recommended sections (adapt as needed): 1. **Problem** (current behavior + pain) 2. **Goals** (explicit outcomes) 3. **Non-goals** (what you’re not doing) 4. **Constraints** (compat, deadlines, security, performance, dependencies) 5. **Proposed design** (data model, APIs, UX flows, before/after behavior) 6. **Risks / mitigations** (what can go wrong, how we’ll detect it) 7. **Tasklist** (phases + numbered tasks with exit criteria)

Tasklist conventions

Put a single, unambiguous execution checklist in the plan:

  • Use **phases** for grouping (Phase 0, Phase 1, …).
  • Use **task IDs** like `0.1`, `0.2`, `1.1`, etc.
  • Each task should be “commit-sized” and include **Exit criteria** that can be verified.
  • Prefer task text that names concrete files/symbols/commands.

Example task shape:

## <N>. Tasklist

### Phase 0 — Foundation
> Goal: <one sentence>

- [ ] **0.1 <Task title>**
  - Deliverables: <files, routes, UI screens, commands>
  - Exit: <tests/commands/manual checks that prove it’s done>

Owner-only decisions

If you hit any unresolved decision that requires product judgment, credentials, or manual local verification:

  • Add an item to `USER_VERIFICATION.md` immediately.
  • Keep the main plan unblocked by marking the decision as deferred to verification.

3) Write `WORKING_MEMORY.md` (live scratchpad)

Purpose: capture what’s currently being worked on, what changed, and what was learned.

Recommended structure:

# Working Memory — <Topic>

> Live scratchpad for agents working through `PLAN.md`. Read before starting work. Update after each task.

## How to use this file
1. Before starting: read Current Task + Key Decisions.
2. When starting a task: set Current Task (task ID + plan).
3. During: log findings/blockers/decisions in Notes.
4. After finishing: move it to Completed Tasks and set the next Current Task.
5. When you discover owner-only decisions/manual checks: add to `USER_VERIFICATION.md`.

## Current Task
**Task:** <id + title>
**Status:** <in progress | blocked | waiting on owner | done>
**Branch:** <branch name>

### Plan
- <bullets>

### Notes
<dated log entries>

## Key Decisions
| Date | Decision | Reasoning |
|------|----------|-----------|
| YYYY-MM-DD | … | … |

## Open Questions / Blockers
- …

## Completed Tasks
| Task | Title | Date | Notes |
|------|-------|------|-------|
| 0.1 | … | YYYY-MM-DD | … |

4) Write `USER_VERIFICATION.md` (owner checklist)

Use this when something is:

  • owner judgment (tradeoffs, “which option”), or
  • manual verification (OAuth flows, API keys, smoke tests, UI polish), or
  • environment-dependent (ports, external services).

Recommended structure:

# Owner Verification — <Topic>

## Open Items
| ID | Item | Why owner-only | Target phase | Status | Notes |
|----|------|----------------|--------------|--------|-------|
| UV-1 | … | … | 2.3 | Open | … |

## Resolved Items
| ID | Decision / Verification | Date | Notes |
|----|--------------------------|------|-------|
| UV-0 | … | YYYY-MM-DD | … |

5) Committing changes in `feature-plans/` (gitignored)

`feature-plans/` is gitignored, so you must force-add updates when committing.

Stage explicitly (no `git add .` / `-A`):

# Non-ignored files stage normally
git add .claude/commands/custom-plan.md

# Feature plans must be force-added
git add -f feature-plans/<topic>

git diff --staged
git commit -m "docs(feature-plans): <topic> plan updates"

For the rest of the session: **commit every time you change anything under `feature-plans/`** using `git add -f feature-plans/<topic>` so the updates don’t get lost.

Read more
Ships withdexto

Agent harness and tookit for building AI agents and agentic applications. CLI and SDKs included

Get the whole plugin

Other commands on dexto.