| name | real-work |
| description | Use when planning a multi-step task or working in plan mode and you need to capture the plan as a durable, resumable artifact — breaking work into phases with per-item checkboxes, completion tracking, autonomous verification, and a handoff summary so a future agent can pick up where you left off. Use when user wants to create or design a plan or mentioned "real work". |
Real Work
Turn planning into a durable, resumable artifact. The plan file — not the
conversation — is the source of truth: it records what to do, what's done, how it
was verified, and how to deploy. Any future agent can resume from it with zero
prior context.
Use when planning multi-step / multi-session work that may outlive the current
session. Skip for trivial single-session tasks.
1. Reach complete understanding first
Do not write the plan until scope is fully understood. Relentlessly ask the
user questions until you both share a complete understanding with no gaps —
treat an unasked question as a future bug.
- Don't stop at the first round; keep going until no ambiguity, assumption, or
open decision remains. Probe edges: scope boundaries (in/out), dependencies,
constraints, success criteria, data, environments, deployment, failure cases.
- Surface every assumption for the user to confirm. If an answer opens a new
unknown, ask the follow-up — drill down recursively.
- Use
AskUserQuestion for concrete choices. When done, summarize the full scope
back and only proceed once the user confirms nothing is missing.
2. Write the plan
Save to plans/<descriptive-name>.md in the repository root (create plans/
if needed). Use this self-documenting template:
# <Work Title>
<1-2 sentence goal and scope.>
## For Future Agents
As work proceeds: mark checkboxes `- [x]` as items complete; when a phase is done,
set its status to `Complete` and write its **Phase Summary** (what was done, key
decisions, anything needed to continue with zero context); run the phase's
**Verification Plan** and record the result before moving on. When all phases are
done, fill in **Final Recap** and **Deployment Plan**.
## Phase 1: <Title>
Status: Not started <!-- Not started | In progress | Complete -->
- [ ] <concrete, actionable item>
- [ ] <concrete, actionable item>
### Verification Plan
- <command/check the agent can run autonomously, with expected result>
### Phase Summary
_(write when phase completes)_
## Phase 2: <Title>
Status: Not started
- [ ] <actionable item>
### Verification Plan
- <autonomous check>
### Phase Summary
_(write when phase completes)_
## Final Recap
_(write when all phases complete: summary of the entire piece of work)_
## Deployment Plan
_(write when all phases complete: step-by-step deployment instructions)_
Common mistakes
- Vague items — each checkbox is a concrete task ("Add retry logic to
PaymentClient.Charge"), not a theme ("improve payments").
- Non-autonomous verification — give runnable commands with expected output,
not "test it manually".
- Wrong location — always the repo-root
plans/ folder.
- Pre-filling summaries — phase summaries, recap, and deployment plan stay as
placeholders until that work actually completes.