Plan capacity

Agent Capacity

Assess agent feasibility for a task using Anthropic's proven decision frameworks. Produce a dispatch plan grounded in official best practices, not guesswork.

Important

• Every classification must trace back to a specific framework from references/decision-frameworks.md. No invented heuristics.
• Report what you find. If a task is straightforward, say so. Do not manufacture complexity to justify the skill's existence.
• Read the actual codebase files a task references before assessing. Armchair scoping without reading code is unreliable.
• Take your time with decomposition. A wrong dispatch plan wastes more time than a careful assessment.

Instructions

Step 1: Identify the Task

Extract the task from one of these sources:

• $ARGUMENTS passed to the skill
• A plan in the current conversation context
• A pasted task description or bug list

Restate the task in one sentence so the user can confirm scope.

Step 2: Decompose into Atomic Units

Break the task into the smallest units where each unit:

• Has a single, verifiable completion condition (test passes, grep confirms change, file exists)
• Can be described in one sentence
• Touches a bounded set of files

List every unit. Number them.

Step 3: Run the Enumeration Test

For the task as a whole, answer: "Can you list all subtasks before starting?"

• YES → This is a workflow problem. Subtasks are predictable. Classify using the six workflow patterns from references/decision-frameworks.md (prompt chaining, routing, parallelization, orchestrator-workers, evaluator-optimizer).
• NO → This needs agent-level autonomy. Identify which specific subtasks are unpredictable and why.

State the answer and reasoning explicitly.

Step 4: Classify Each Unit

Apply Anthropic's decision ladder to each atomic unit. Consult references/decision-frameworks.md for the full framework.

For each unit, assign exactly one classification:

Classification	Criteria	Claude Code approach
Inline	Single prompt, 1-2 files, clear completion condition	Do it directly in main conversation
Explore subagent	Read-only investigation, codebase search, file discovery	subagent_type: Explore
General subagent	Self-contained task, clear deliverable, 3-10 tool calls	subagent_type: general-purpose
Sequential chain	Depends on a previous unit's output	Must run after its dependency completes
Human checkpoint	Ambiguous requirement, judgment call, or moving-target risk	Pause and ask the user

Step 5: Check Delegation Specs

For every unit classified as a subagent, verify all four elements of Anthropic's delegation spec are present:

1. Objective — Is the goal specific and unambiguous?
2. Output format — Will the agent know what to return?
3. Tool guidance — Does it know which files/tools to use?
4. Task boundaries — Are scope limits explicit?

Flag any unit missing an element. A subagent dispatched without all four will "duplicate work, leave gaps, or fail to find necessary information."

Step 6: Flag Failure Patterns

Scan the full task for known failure patterns from references/decision-frameworks.md:

• Error accumulation — Long chains where each step's error compounds
• Over-delegation — Units that a single grep or one-file edit would handle faster than a subagent
• Shared-context splitting — Phases that need shared context being split across subagents
• Moving targets — Fixing unit A changes the landscape for unit B
• Vague delegation — Units missing delegation spec elements (caught in Step 5)

Step 7: Output the Dispatch Plan

AGENT CAPACITY ASSESSMENT
═══════════════════════════════════════════
TASK: [one-line summary]
UNITS: [N total]
ENUMERABLE: [YES/NO — from Step 3]

─── DISPATCH PLAN ──────────────────────────

Phase 1 — Parallel [units that can run simultaneously]
  [U1] [INLINE] [one-line description]
  [U2] [SUBAGENT:explore] [one-line description]
  [U3] [SUBAGENT:general] [one-line description]

Phase 2 — Sequential [units depending on Phase 1]
  [U4] [INLINE] [one-line description] (depends on U1)

Phase 3 — Human checkpoint
  [U5] [HUMAN] [what needs user judgment and why]

─── FAILURE RISKS ──────────────────────────
[List any flagged patterns from Step 6, or "None identified"]

─── DELEGATION SPECS ───────────────────────
[For each subagent unit, show the 4-element spec or flag what's missing]

═══════════════════════════════════════════
TIER: [SIMPLE / COMPARISON / COMPLEX — from the three-tier scale]

Tier definitions (from Anthropic's multi-agent research system):

• SIMPLE — 1 agent, 3-10 tool calls. Most single-file edits and lookups.
• COMPARISON — 2-4 parallel subagents, 10-15 calls each. Module-level analysis.
• COMPLEX — 10+ subagents with divided responsibilities. Cross-codebase work.

Step 8: Offer Execution Options

After the report, offer exactly 2 options:

1. "Execute this dispatch plan?" — Proceed to run the phases as described.
2. "Adjust the plan?" — User modifies scope, grouping, or classifications.

Error Handling

1. Task is too vague to decompose: Report: "Task lacks sufficient detail. Specify: [what's missing — files, goals, success criteria]." Do not guess at decomposition.
2. Task is a single atomic unit: Report it as SIMPLE tier, INLINE classification. Do not over-decompose. A one-line fix does not need phases.
3. Task references files that don't exist: Flag as a potential issue — the task may be creating them, or it may be based on stale information.
4. No plan in conversation context: If invoked without $ARGUMENTS and no plan exists, say: "No task found. Pass a task description: /agent-capacity [description] or create a plan first."
5. All units are inline: This is valid. Report SIMPLE tier and note that no subagents are needed. Do not manufacture reasons to use subagents.

Examples

Example 1: Simple bug fix

Input: "Fix the risk.py per-market cost check — it doesn't include existing position cost"

TASK: Fix risk.py per-market cost validation to include existing position cost
UNITS: 1
ENUMERABLE: YES

Phase 1 — Inline
  [U1] [INLINE] Add existing position cost to per-market check in risk.py

FAILURE RISKS: None identified
TIER: SIMPLE

Example 2: Multi-item audit list

Input: The 12-item mm-alpha audit list (bugs + dead code + strategy gaps)

TASK: Address 12 audit findings across mm-alpha codebase
UNITS: 12
ENUMERABLE: YES

Phase 1 — Parallel (independent fixes)
  [U1] [INLINE] Fix risk.py per-market cost check (1 file, 1 line)
  [U8] [SUBAGENT:general] Remove 4 dead code items (_ensure_sdk_client, get_skew_cents, check_inventory_limits, get_expiry_bonus)
  [U5] [SUBAGENT:explore] Investigate max_quote_size usage — confirm it's truly unused

Phase 2 — Sequential
  [U2] [INLINE] Add pagination to get_fills() (depends on understanding current implementation)
  [U6] [HUMAN] Amend threshold inconsistency — which behavior is correct? (needs user judgment)
  [U7] [INLINE] Add market dequeue logic for out-of-zone markets

Phase 3 — Design required
  [U3] [HUMAN] Adverse-selection detection — needs design approval before implementation
  [U4] [HUMAN] TIME_OF_DAY_FILTERS integration — needs strategy decision

FAILURE RISKS:
- Moving target: if dead code removal (U8) changes imports, U2/U7 may need adjustment
- Over-delegation risk: U1 is a one-line fix, do not subagent it

TIER: COMPARISON