Back to skills
SKILL.md
Writing Plans
ASecurityWrite file-level, bite-sized implementation plans before coding. Use when you have a spec or multi-step requirements and need an actionable plan before touching implementation.
- 2 stars
- 0 votes
- 0 copies
- 0 views
- Added September 23, 2026
Security analysis
100/100npx -y skills add danmo-ai/danmo-work --skill writing-plans --agent claude-codeAre you the author of Writing Plans?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/danmo-ai-writing-plans)---
name: writing-plans
source: builtin
description: Write file-level, bite-sized implementation plans before coding. Use when you have a spec or multi-step requirements and need an actionable plan before touching implementation.
license: MIT
compatibility: Requires read_file, grep, glob; write only on Default/Team when the user asks to save the plan file
metadata:
author: danmo-work
version: "1.0"
category: coding
adapted_from: "https://github.com/obra/superpowers/tree/main/skills/writing-plans"
upstream_license: MIT
---
# Writing Plans
> Adapted from obra/superpowers `writing-plans` (© Jesse Vincent / contributors, MIT).
> Rewritten for Danmo Work tools and agents; not a verbatim copy.
Write plans that a skilled engineer with little repo context can execute: exact files, concrete steps, tests, and commands. Prefer DRY, YAGNI, TDD, frequent commits.
## Workflow
1. **Explore** — map relevant files with `glob` / `grep` / `read_file`. Follow existing patterns.
2. **Scope check** — if the work spans independent subsystems, propose separate plans (each delivers testable value).
3. **File map** — list create/modify/test paths and each file’s responsibility before tasks.
4. **Tasks** — bite-sized units with their own test cycle; each ends in an independently verifiable deliverable.
5. **Self-review** — spec coverage, no placeholders, consistent names/types across tasks.
6. **Handoff**
- **Planner (read-only):** always deliver the complete plan in the assistant reply (session event stream). Do not call `write` or claim a file was saved.
- **Default / Team:** deliver in chat by default; only `write` a plan file when the user explicitly asks to save it. Then ask whether to implement.
## Plan header (required)
```markdown
# [Feature] Implementation Plan
**Goal:** one sentence
**Architecture:** 2–3 sentences
**Tech stack:** key pieces
**Constraints:** version floors, naming, platforms (verbatim from spec)
```
## Task template
```markdown
### Task N: [Name]
**Files:**
- Create: `path`
- Modify: `path` (why)
- Test: `path`
**Steps:**
- [ ] Write failing test (show code or exact assertion)
- [ ] Run test — expect FAIL for reason X (exact command)
- [ ] Minimal implementation (show code or precise edit)
- [ ] Run test — expect PASS (exact command)
- [ ] Commit message suggestion
```
## No placeholders
Never leave: TBD/TODO, “add appropriate validation”, “write tests for the above” without actual tests, “similar to Task N”, or steps without exact paths/commands.
## Anti-patterns
- Vague “update the service layer” without files
- Giant tasks that mix unrelated deliverables
- Plans that assume unspoken domain knowledge
- Skipping self-review for type/name drift across tasks
## Stop condition
Deliver the complete plan in chat (and path only if Default/Team saved a file). Offer to implement next with Single-Agent or Multi-Agent mode; do not start coding inside this skill unless the user already approved execution.
Attribution
Comments
Loading comments…
Newsletter
Get the weekly list of safe new skills, and the ones we pulled.