On-demand /consult skill — a Claude Fable 5 plan-consult plus an automatic warm diff-review on any task. Opus drives; Fable consults.
Agent
Implements a specific, already-diagnosed code change (bug fix, small improvement, dependency cleanup). Use PROACTIVELY whenever a concrete code modification has been identified and approved. Requires a precise spec - file, function, expected change, and verification steps.
Agent
Executes an approved structural refactor or multi-file feature - extracting functions, consolidating duplication, improving performance of a specific hot path, reorganizing a module, building a spec'd feature. Use when the change spans multiple functions or files but has been scoped and approved. NOT for one-line bug…
Agent
Sonnet read-only discovery. Delegate when discovery needs ≥3 searches or spans unfamiliar territory. Give it the question plus starting paths; it returns a conclusion, absolute paths, and minimal excerpts — never file dumps.
Agent
Fable 5 planner for the /fable skill. Invoked once by /fable with a plan brief — intent, task verbatim, session constraints, file map, pasted load-bearing code, and 1–3 open questions — and returns the plan: numbered steps with risks, checkpoints, assumptions. Resumed at task end for the warm diff-review…
Agent
Reviews uncommitted changes or a recent commit for correctness, regressions, and spec compliance. Use PROACTIVELY after coder or engineer completes work, before the orchestrator accepts it. Read-only - never modifies files.
Agent
Drives the REAL running application end-to-end after a deploy to prove a change works live - not just in unit tests. Give it a concrete scenario ("trigger X, assert Y appears within Z"), how to launch/reach the app, and how to clean up. NOT for code changes (use coder), NOT for running test suites (use verification).
Agent
Writes tests - characterization tests to pin down existing behavior before a refactor, regression tests for a fixed bug, or coverage for untested critical paths. Use PROACTIVELY before any risky change to untested code.
Agent
Haiku test/lint/build runner. Delegate any verification whose raw output the driver would otherwise scroll inline. Give it exact commands plus what "pass" means; it returns pass/fail and distilled failures.