agentsclimarketplace

Phobos plan

Skill broisnischal/phobos/skills/phobos-plan

Analyze the user's request before acting — extract every ask, resolve ambiguity in one batch, order work by dependency and risk, and decide what to do first. Use when a request is multi-part, vague, or large; when asks conflict; or when the user says "phobos-plan", "plan this", "prioritize", or "what first".From its SKILL.md

Install
npx -y skills add broisnischal/phobos --skill phobos-plan

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

SKILL.md

3.0 KB, as published. Nobody here has run it

phobos-plan

Wrong order and misread intent are the two most expensive mistakes — each one costs a full redo. Spend 30 seconds parsing before spending 30 minutes building.

1. Parse the request

Extract into four buckets (mentally for small tasks, written for large ones):

  • Explicit asks — every deliverable actually stated. Multi-part messages hide asks in the middle; count them, address all.
  • Implicit needs — what the ask requires but doesn't state (a new endpoint implies auth + validation; "make it faster" implies measuring first).
  • Constraints — stack, style, "don't touch X", deadlines, things the user already decided. Never re-litigate a decision the user has made.
  • Unknowns — anything where a wrong guess forces a redo.

The user describing a problem or thinking out loud ≠ requesting a change. Deliverable there is your assessment — report, don't fix.

2. Resolve unknowns — once

  • Changes the outcome → ask. Batch every question into one message; drip-fed questions are the #1 source of wasted round-trips.
  • Doesn't change the outcome → pick the sensible default, state it in one line, proceed.
  • Answerable from the code/repo → answer it yourself; never ask the user what a grep can tell you.

3. Order the work

Priority, first match wins:

  1. Blockers — anything other items depend on (schema before queries, API shape before UI).
  2. Cheap decision-changers — a 2-minute check that could invalidate the plan (does the lib support X? does the bug even reproduce?). Do these before committing to a direction.
  3. Risk — the part most likely to fail or surprise. Fail fast while the context is small.
  4. User-visible value — what the user actually asked for, core before edges.
  5. Polish — cleanup, docs, nice-to-haves. Last, and only if asked or trivial.

Tie-breaker: shortest task first — it unblocks feedback soonest.

4. Emit the plan — sized to the task

  • 1–2 steps → no plan, just do it. A plan longer than the task is slop.
  • 3+ steps or any reordering of the user's own order → 3–7 numbered lines: what, in what order, one clause on why the order matters. Call out any dropped or deferred item in a plain sentence — never silently drop an ask (no warning-symbol prefix).
  • Then execute immediately. A plan without the first step taken is a stall.

Anti-patterns

  • Asking "should I proceed?" on work that follows from the request.
  • Planning in prose what could be discovered faster by reading the code.
  • Re-planning mid-task without new information.
  • Treating the user's mention-order as priority-order — dependency beats position.

Controls: "stop phobos-plan" / "normal mode" → off until re-invoked.

Keep looking

Skills are one crate of 627,801. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.