Skip to content
SHRINESHRINE

Delegation Fit

Ratified

Some tasks cost more to review than to write.

The Pattern

Section titled “The Pattern”
  • Before prompting, decide per task: delegate, pair, or write by hand
  • Keep decisions that belong to a human with the human; decide first, then delegate the rest
  • Brief delegated work in one message with explicit non-goals

When to Use

Section titled “When to Use”
  • Everyday interactive work, before the first prompt
  • A feature with parts of different difficulty
  • Output from past tasks of this kind needed heavy correction

When Not to Use

Section titled “When Not to Use”
  • Multi-hour or unattended runs; use Spec, Then Build
  • Pure exploration, where the model’s interpretation is the point

Fit Signals

Section titled “Fit Signals”

For delegating

  • The outcome is checkable by a command
  • The repo has an exemplar to mirror
  • Boilerplate, broad search, or mechanical change
  • Reversible

Against delegating

  • You already hold the full solution and typing is the only cost
  • No check exists, so you become the check (Verification Loops)
  • Tacit design judgment that is hard to write down
  • Review would take longer than writing

These signals are judgment calls and shift as models change. Reassess them on the Deliberate Currency cadence.

Keep With the Human

Section titled “Keep With the Human”
  • Architecture choices
  • Ambiguous requirements
  • Anything Human in the Loop lists under low confidence
  • Decide these first; then delegate the work they unblock

Middle Mode: Pair

Section titled “Middle Mode: Pair”
  • The model explores, drafts tests, or lists edge cases
  • The human writes the core logic
  • Useful when the core needs tacit judgment but the edges are mechanical

Sizing

Section titled “Sizing”

Brief Shape

Section titled “Brief Shape”

Extends the problem prompt in Problem Before Prescription for small delegated tasks. It is a small-task subset of the Spec Fields in Spec, Then Build.

  • Goal: the problem, not the fix
  • Constraints: what must not change
  • Non-goals: what not to touch or start (the spec’s “Out of scope”)
  • Mirror: an existing file or function to follow
  • Done: the command that proves it
  • Ask: state assumptions or ask questions before editing

Prior art: Claude Code best practices recommend pointing the model at an existing pattern in the codebase to follow.

Worked Example

Section titled “Worked Example”
  • Illustrative case (composite, not a transcript)
  • Feature: add CSV export to the orders report

Split

Part Mode Why
Export format and column contract with the finance team By hand Ambiguous requirement; a human decision
Streaming large result sets without loading all rows Pair Model drafts edge-case tests; human writes the streaming core
Handler, route, and auth wiring Delegate GET /reports/{id}/pdf is an exemplar; tests check it

Vague prompt

  • “Add CSV export to the orders report”
  • Outcome: one large diff; invented columns, a new CSV library; five correction turns

Brief

  • Goal: finance needs the orders report as CSV
  • Constraints: existing report API unchanged; columns exactly as in docs/orders-export.md
  • Non-goals: no streaming changes; no new dependencies
  • Mirror: follow GET /reports/{id}/pdf for route, auth, and tests
  • Done: make test passes, including new test_orders_csv_columns
  • Ask: list assumptions before editing
  • Outcome: one question about date format, answered; one correction, a header casing
  • Minutes (illustrative): both rows delegate the same slice, the CSV route, column mapping, and its test; by hand is that same slice typed directly
Path Minutes Against 30 by hand
Vague prompt 1 prompt + 44 review and correction = 45 15 worse
Brief 7 brief + 9 review = 16 14 better

Anti-patterns

Section titled “Anti-patterns”
  • Delegating the architecture decision, then correcting the architecture in review
  • Delegating work you could type faster than you could brief
  • A brief with no non-goals, so the diff grows past the task
  • Fit decided once and never revisited as models change
Section titled “Related”