plan - SKILL.md Agent Skill

name: plan description: Break an approved spec or design into a precise, executable implementation plan. Use after the what/why and the shape are agreed, and before writing implementation code for any multi-step task.

██████╗ ██╗      █████╗ ███╗   ██╗
██╔══██╗██║     ██╔══██╗████╗  ██║
██████╔╝██║     ███████║██╔██╗ ██║
██╔═══╝ ██║     ██╔══██║██║╚██╗██║
██║     ███████╗██║  ██║██║ ╚████║
╚═╝     ╚══════╝╚═╝  ╚═╝╚═╝  ╚═══╝

Plan in vertical, executable slices

A plan good enough that a context-free executor with questionable taste could follow it without guessing. Vague plans get vague work.

Slices are vertical tracer bullets

Each task cuts end-to-end through whatever layers the artifact actually has and leaves the program working. For a web feature that may be schema → API → logic → UI → tests; for a CLI it's flag → handler → output; for a library it's public function → behavior → test. The rule is end-to-end-to-an-observable-result — not a fixed layer list. Never slice horizontally (all of one layer, then all of the next) — horizontal slices can't be verified until the end and hide integration problems until they're expensive.

Among valid first slices, pick the one that retires the most risk — the integration most likely to be wrong (the unproven third-party call, the gnarly migration), not merely the easiest. De-risk first; polish later.

Each task is bite-sized and test-shaped

Size each task to one RED→GREEN→commit cycle — the minimum change that makes exactly one new test pass. If a task needs two assertions about two different behaviors, split it. (The loop itself lives in forge:tdd — don't restate it; plan just chooses the slices.)

Write each task so a context-free executor can run it. Give every task these fields:

id (T1, T2…) + a one-line intent
target — the file(s) and the function/signature it adds or changes
test — the assertion that should fail first, then pass (the success criterion — a concrete check, never "make it work")
blocked-by — real task IDs, or "none"

(This is forge:principles #4 made concrete.)

No placeholders

The executor may read tasks out of order and has no context beyond the plan. Therefore:

No "TBD", no "add appropriate error handling", no "handle edge cases."
No "similar to Task 3" — restate the actual intent/code, because Task 3 may not have been read.
If a value, path, or signature matters, write it down.

Order and parallelism

Order by dependency. Reference blockers by real task IDs ("blocked by T2"), added only once T2 exists.
Mark which tasks are independent and safe to parallelize — hand those to forge:orchestrate to run concurrently with fresh agents.
Keep the plan in a file, not just the conversation — it's a durable artifact that survives compaction.

Exit

Execute with forge:tdd (one slice at a time), request forge:review after meaningful chunks, and forge:verify before claiming done.