<< All versions

Skill v1.0.0

currentLLM-judged scan90/100
szokeasaurusrex/pi-agent/delegation-plan
──Details
PublishedOctober 21, 2026 at 07:37 PM
Content Hashsha256:b3c09cc063ddee21...
Git SHA6ed8b7bbf5dc
──Files
Files (1 file, 2.9 KB)
SKILL.md2.9 KBactive
SKILL.md · 100 lines · 2.9 KB

version: "1.0.0" name: delegation-plan description: Write plans intended for a cheaper or less capable implementation model. Use when the user asks for a plan suitable for another LLM, junior engineer, intern, cheaper model, coding agent, or implementation delegate.


Delegation Planning

Use this skill to write implementation plans for a capable but lower-reasoning executor.

Goal

Write a plan that removes ambiguity without doing the implementation reasoning that the delegate can do cheaply.

Think like a senior engineer and prompt engineer:

  • Specify the objective, constraints, and success criteria.
  • Define clear boundaries and decision rules.
  • Leave routine code navigation and implementation details to the delegate.

Planning principles

  • Do not paste full implementations unless the exact code is the requirement.
  • Do not provide large code blocks that solve the task.
  • Prefer descriptions of required behavior over exact replacement snippets.
  • Give file/module targets, not line-by-line edits, unless the location is genuinely ambiguous.
  • State what must not change.
  • State how to handle known edge cases.
  • Include validation commands and expected outcomes.
  • Include escalation criteria: when the delegate should stop and ask.

Required structure

Use this structure unless the user requests another format:

markdown
# <Task title>
## Objective
<One paragraph describing the user-visible behavior to achieve.>
## Scope
-In scope: <files, modules, behavior>
-Out of scope: <explicit non-goals>
## Constraints
-<compatibility, minimality, performance, style, or API constraints>
## Implementation guidance
1.<Concrete step with target files/modules and intended behavior.>
2.<Concrete step.>
3.<Concrete step.>
## Edge cases
-<case>: <required handling>
## Tests
-<test to add/update and what it should prove>
## Validation
Run:
-`<command>`
Expected result:
-<what passing looks like>
## Stop and ask if
-<condition that requires human/senior decision>

Specificity calibration

Use more detail for:

  • Non-obvious invariants.
  • Risky migrations.
  • Security, data loss, auth, or compatibility concerns.
  • Places where multiple valid implementations would conflict with user intent.

Use less detail for:

  • Obvious syntax.
  • Mechanical refactors.
  • Test fixture boilerplate.
  • Exact code the delegate can infer safely from surrounding code.

Anti-patterns

Avoid:

  • Writing exact helper functions when a behavior description is enough.
  • Replacing the delegate's design work with full code.
  • Hiding choices as assumptions.
  • Saying "update tests" without naming the behavior each test must verify.
  • Over-constraining harmless implementation details.

Final self-check

Before finalizing, verify:

  • A motivated intern could start without guessing the goal.
  • The delegate still has meaningful implementation work to do.
  • Every ambiguity that affects correctness has a rule or escalation path.
  • The plan says how to prove the change works.
All versions