| name | long-horizon-task-execute |
| description | Execute a complex task step-by-step using the confirmed design document. Load this Skill after the user has confirmed the design document produced by the long-horizon-task-design Skill. Input is the design document path. |
| metadata | {"short-description":"Execute complex task step-by-step per design document"} |
Long-Horizon Task Execution
This Skill drives the step-by-step execution of a complex task whose design document has been confirmed by the user. You receive the design document path as input. Work through the steps below in order.
Step 1 — Read the Design Document and Extract Steps
Read the full design document at the provided path. Locate the "Implementation Steps" section and extract every step with its current status.
Step status is indicated by the prefix on the Status field:
| Prefix | Meaning |
|---|
[ ] | Not started |
[x] | Done |
Step 2 — Display the Execution Plan
Before executing anything, show the user the full step list in this format:
## Execution Plan
[ ] Step 1: <title>
[ ] Step 2: <title>
[x] Step 3: <title> (done)
...
Progress: <N done> / <total> completed
Map each step's current status prefix to the display above. Do not wait for user confirmation — begin executing immediately after displaying.
Resuming an interrupted task: if some steps are already marked [x], display the current state and explain which step you are resuming from.
Step 3 — Execute Steps in a Loop
Repeat the following cycle for each step that is not yet done ([ ]), in document order:
3.1 Decide execution mode
Assess whether this step is complex enough to run in a sub-agent:
- Use
run_subtask if the step involves significant exploration, cross-module changes, or multiple interdependent edits that benefit from an isolated context window.
- Execute directly if the step is straightforward — e.g., a single-file edit, a configuration change, or a simple addition with no unknowns.
No hard threshold applies; use your judgment. State your decision in one line before proceeding.
3.2 Direct execution path
If you chose to execute directly, complete the step yourself using available tools. When done, proceed to 3.5.
3.3 run_subtask path — pre-execution research
Before assembling the instruction, read the source files and design document sections relevant to this step. Produce a brief research summary covering
3.4 run_subtask path — assemble and call
Use this template to build the instruction argument:
Context:
- Working directory: <absolute path of the current working directory>
- Design document: <absolute path to the design document>
Research summary:
<findings from 3.3: key patterns, interfaces, dependencies, constraints>
Previous step result:
<the structured summary written back from the most recently completed step>
(If this is the first step, write: "N/A — this is the first step.")
Your task:
Implement step <N>: <step title>
Call run_subtask with this instruction string.
3.5 Handle the result
Treat the outcome as either success or failure:
Success (direct execution completed, or run_subtask returned status == "completed")
- In the design document, change the step's Status prefix from
[ ] to [x]
- Append a structured "Result" sub-field on the step entry: list files changed and key decisions made
- Save the document
- Display the updated Execution Plan (same format as Step 2)
- Continue to the next not-started step
Failure (direct execution failed, or run_subtask returned "failed" or "blocked")
- The step status in the document remains
[ ] — do not modify it
- Report to the user: the step title, what went wrong, and (if from
run_subtask) the blockers content
- Ask the user for instructions
- Wait for the user's response before proceeding
Step 4 — Finish
When all steps are marked [x], output a completion summary to the user:
Task complete.
Steps executed: <N>
<brief summary of what was accomplished overall>
