AI Fluency Delegation Plan

A Claude skill that turns a batch of open issues or a project phase into a concrete delegation plan: per work item, who leads (human or AI), in what interaction mode, and where human judgment should concentrate.

Built on Anthropic's AI Fluency framework — the 4Ds: Delegation, Description, Discernment, Diligence — applied to real tracker state instead of summarized as theory.

What it produces

A phase-scoped markdown document saved to docs/ (e.g. delegation-plan-implementation.md) containing:

• Problem awareness — the destination, definition of done, and the phase's main failure mode
• Platform awareness — what makes this batch delegable, where AI is strong, and where risk concentrates
• Task delegation table — per item: who leads / interaction mode / named human-discernment focus, with rationale for the non-obvious calls
• Cross-cutting Description + Diligence — the standardized per-item process, how performance is recorded, and the non-delegable sign-offs listed by name

Interaction modes use the framework's vocabulary only:

Mode	Meaning
Automation	AI performs the task, human reviews the output
Augmentation	Human and AI collaborate, human in the loop
Agency	AI works independently within human-set guardrails

When to use it

• "Generate a delegation plan for the open issues"
• Entering a new phase (spec → build, build → release) where the human/AI split should be re-decided
• Any time you invoke the AI Fluency framework / 4Ds for planning human–AI collaboration

Not for one-off "should I prompt the AI to do this?" decisions — those are decided inline.

How it works

The skill refuses to plan from memory or the prompt alone. It grounds every plan in three sources first:

1. Live tracker state — issue numbers, titles, labels, and blocking edges (e.g. gh issue list). The blocking graph gives work order; labels like ready-for-agent already encode delegation signal.
2. Fixed points — the spec/PRD and normative docs each item is verified against. What makes work delegable is precisely that these exist.
3. Prior delegation plans — plans are phase-scoped and iterative. Each plan links its predecessor, honors what it promised, and states what triggers the next one.

Principles baked in

• Focused review beats uniform review — reading every diff line of N tickets catches nothing; each automation item names ONE thing worth human eyes.
• Adversarial review for silent failures — secrets, concurrency, destructive commands: green tests can hide bad outcomes, so the human attacks rather than reads.
• The yardstick stays human-led — AI must not both implement a system and declare its own acceptance passed.
• Pull external risk forward — items with real auth, real costs, or undocumented dependencies go to the front of the frontier while there's still slack.

Installation

Copy the skill folder into your Claude skills directory:

skills/
└── ai-fluency-delegation-plan/
    └── SKILL.md

Works in Claude Code, Cowork, and any environment that loads Agent Skills.

Usage

Generate a delegation plan for the open issues in this repo

or

We're moving from spec to build — create a Phase 2 delegation plan using the 4Ds

The output lands in docs/ with a kebab-case, phase-distinct filename. Prior phases' plans are never overwritten — each is the record of that phase's arrangement.

Mistakes the skill guards against

• Inventing interaction-mode tiers ("Full/Partial") instead of automation / augmentation / agency
• Skipping problem and platform awareness and jumping straight to a task table
• Planning from conversation memory instead of live tracker state and blocking edges
• Missing phase linkage — no predecessor plan, no trigger for the next one
• Uniform review depth, or an unnamed "human reviews" with no focus

Reference

• AI Fluency: Framework & Foundations — Anthropic's free course on the 4Ds