write-a-goal

Turn a rough ambition into a measurable, REACHABLE goal — a capability statement, a precisely-named metric, ≥2 numeric evals, an acceptance rule, and non-goals — with a feasibility check that catches

SKILL.md

GitHub

/write-a-goal

Usage: /write-a-goal <the ambition, or what "good" looks like>

A goal you can't measure is a wish. A goal you can't reach is a trap. A goal a degenerate output can fake is a lie. This skill turns an ambition into a goal that /goal can autonomously pursue and a fresh-eyes agent can actually certify — and it refuses to hand you a target that's impossible or game-able.

The shape a good goal must clear

  1. Capability statement — one sentence: "Given X, the system does Y." Concrete, observable.
  2. The exact metric — the single thing measured, named precisely enough that two people would compute the same number (see Check A).
  3. ≥2 numeric evals — real test cases, each with a target number and how it's scored. One eval is an anecdote; two+ is a bar.
  4. Acceptance rule — when is it done? (e.g. "every eval ≥ target on a fresh run, confirmed by a fresh-eyes agent.")
  5. Non-goals — what this explicitly is not, so the loop doesn't gold-plate the wrong thing.

The two checks that make or break it (hard-won)

Check A — Pin the exact metric (kill vague percentages)

"Make it 99% similar" is meaningless until you answer similar HOW. The same artifact can score 20% or 95% depending on the measure. Force the choice:

  • Name the measure (e.g. visual SSIM on edge maps, not "looks the same"; structural token overlap, not "matches the design"; p95 latency, not "fast").
  • State how it's computed and what tool reports it.
  • If one number hides several things, split it (e.g. "fidelity" → visual + structural + copy).
  • Robustness clause: require the metric to resist gaming. Ask "what's the dumbest output that could score high?" — if a blank/empty/degenerate result can fake the target, the metric is broken; add a guard before you trust it.

Check B — Feasibility / ceiling check (don't set a trap)

Before committing the target, pressure-test it against what the architecture can actually reach:

  • What's the theoretical ceiling for this approach? (e.g. a small local model reconstructing from a text description can't pixel-match a real site — the inputs it's missing cap it far below 100%.)
  • Is the target inside that ceiling, with headroom? If not, say so plainly and propose a reachable target instead — never accept an impossible number silently.
  • Name the lever that would raise the ceiling (a different input, a different model class) vs. the levers that just optimize within it — so effort goes where it pays.

What you do as the assistant when this is invoked

  1. Extract the ambition — what does the user actually want, and what would make them say "yes, that's it"? Ask at most one clarifying question if it's truly ambiguous.
  2. Draft the 5-part goal (capability · metric · ≥2 evals · acceptance · non-goals).
  3. Run Check A — pin the exact metric; split vague numbers; add the robustness guard.
  4. Run Check B — feasibility. If the target is unreachable or game-able, lead with that, and offer a reachable, robust alternative target (recommended-first).
  5. Output the goal spec as a small artifact (capability, metric+tool, evals with numbers, acceptance, non-goals, and the feasibility verdict).
  6. Offer to hand off to /goal to pursue it autonomously, or /full-cycle if it's a build.

Anti-patterns — real cases this skill exists to prevent

  • Vague + impossible: "Replicate this homepage to 99%." Two failures at once — the metric is undefined (visual? structural? pixel?), and the target is unreachable for the architecture (a local 14B reconstructing from a text digest plateaus ~30% visual; placeholder-vs-real images alone cap it). The fix: define the measure (visual SSIM with a blank-guard), and set a reachable bar (e.g. structural ≥ 85% + visual ≥ 70% with real images), naming "a vision model" as the lever that would raise the ceiling.
  • Game-able metric: a blank page scoring 98% because it matched the site's dark background. A target with no robustness guard rewards the dumbest output.
  • Single anecdote as a bar: one passing example ≠ a goal. Require ≥2 evals.
  • A vibe with no number: "make it feel premium." Either find a measurable proxy or admit it's a taste call, not a goal.

Companion: meta/learning-loop-rubric.md (scoring what to promote) and meta/always-recommend.md (lead with the recommendation — including when proposing a reachable target). /goal pursues a goal; /write-a-goal forges one worth pursuing.