Work Plan

Work / 3. Plan

Produce an implementation-ready plan without changing source files.

Make the plan protect reliability, restraint, modular boundaries, the simplest complete implementation, human authorship, clean breaks over obsolete compatibility, meaningful tests, and proof over status reports. Prefer existing dependencies and durable architecture over speculative packages or stopgaps.

Fixed route

• If the exact subagent model, effort, or goal tools are unavailable, return blocked with the missing capability. Do not substitute a model or perform the analyst's work in the parent.
• Launch exactly one gpt-5.6-sol subagent at xhigh effort with fork_turns: none.
• Give it a fun call sign, the exact task statement, source of truth, read-only ownership, constraints, acceptance checks, and validation requirements.
• Require it to inspect independently and return concrete evidence, not a status report or a claim that an unproved step is “routine.”
• Require it to call create_goal with the exact analysis goal before inspection and update_goal complete only after the required evidence is collected.
• Let the analyst run until it returns the plan, a real blocker, or a user-input request. Do not cancel it because a polling window or arbitrary wall-clock interval elapsed; a wait timeout is not a failure.
• The analyst must not edit, commit, deploy, spawn, or delegate.
• Scale depth to the task. Inspect only relevant paths, direct consumers, and evidence needed for the handoff; return the plan without process narration and stop when every required field is supported.
• Report the successful launch as a compact table with Agent, Working on, Goal, Ownership, and Model.

Build the plan

The parent defines the mission. The analyst inspects only relevant context, traces current behavior through real consumers, and resolves evidence-backed decisions. Keep early alternatives independent; do not let the first elegant reduction become the plan without checking it against the exact task. The parent verifies paths, symbols, commands, and assumptions before returning the handoff.

For test work, inspect the implementation, existing tests, public APIs, and actual user flows before proposing a test. Put the four-line test plan before any test implementation assignment, and name tests deliberately omitted when no real behavior justifies them.

Confirm that implementation serves a defined user outcome rather than merely being cheap to generate. If the plan includes prototypes or variants, require each to answer a named question and include the evidence, selection rule, or kill criterion that determines what survives.

For qualitative decisions, record the fundamentals, references, context, and tradeoffs that make the chosen direction appropriate. Do not use preference or trend-following as the entire rationale.

## Work / 3. Plan

**Mission:** 

**Source of truth:** 

**Current behavior and evidence**
- 

**Done when**
- 

**Does not count**
- 

**Decisions, constraints, and non-goals**
- 

**Taste rationale**
- 

**Alternatives rejected**
- 

### Implementation map
| Step | File / symbol | Exact change | Reason / invariant |
|---|---|---|---|
| 1 |  |  |  |

**Executor assignments:** 

### Validation
| Command or flow | Expected evidence | Proves |
|---|---|---|
|  |  |  |

**Testing**
- **Behavior being protected:** 
- **Why it matters:** 
- **Test level:** 
- **Exact observable assertion:** 
- **Tests deliberately not added:** 

**Risks or material unknowns**
- 

**Ready:** Yes | No — 

Assume the plan may be executed by a lower-capability model with no access to hidden context. Make it self-contained: name exact files and symbols, ordered edits, patterns to follow, invariants, edge and failure cases, acceptance evidence, and commands with expected results. Resolve judgment-heavy decisions in the plan instead of delegating them to the executor.

A plan is not ready when it merely restates the task, reduces it to an unproved assumption, omits a required acceptance check, or reports progress without evidence. If a route is blocked, state the exact gap and propose another round only for a materially new mechanism. Do not launch an executor or edit files from this skill. Hand a verified ready plan to $work for execution.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: xcaeser
• Source: xcaeser/work-skill
• License: MIT
• Homepage: https://skills.sh/xcaeser/work-skill

Install and usage instructions live in the source repository linked above.