clarity & grace · a research‑writing skill
Clarity & Grace is not a prompt. It is a master who holds the plan, a writer, a reviewer who reads cold, and blind judges who decide what stays, with you as the author. Built from one trusted book on clear writing.
From the skillfed.io lab · found by any agent through Skill Federation
what a skill really is
One file of advice. One model follows it: writes a draft, reads it back, approves it. It reads what it meant, so it can’t see what a reader pays. It polishes sentences and keeps the structure, and you rewrite the paper by hand.
Give the writing to one agent and the reading to another. The reviewer reads cold, with only the text and who the reader is. It names the habit behind the problem; the actor changes the habit, not the cited sentence. The file splits with them: how to act, how to read.
From inside a section nobody can see the paper’s shape. A master holds the plan at a bird’s-eye view and changes structure only through it. Blind judges compare every new version with the last, in both reading orders, so progress is measured, not asserted.
Who owns what, written down: the plan to the master, everything inside a section to the writer, the reading to the reviewer, the verdict to the judges. The master never sets wording and never writes a more detailed brief. Nobody reaches into a layer they don’t own.
You set the intent, clear the plan and read the result. Your reading is final. What you catch that the agents missed is traced to the layer it came from and fixed there, so it never comes back.
A collective of well-tuned agents working on your task semi-autonomously, with you at the top. The SKILL.md you pictured is still here: one box in the corner. The skill is everything around it.
the secret sauce
Three loops run behind every skill from the lab: the evaluators are tested, the skill is trained on real work, and it tunes to you.
Reviewers and judges get a test set of their own: versions deliberately broken in ways a reader would feel. They have to catch every one. Each blind spot a user finds becomes a new case.
fractured · shuffled · figures removed · a frame repeated · plus whatever the next user catches
The skill runs on a real paper. What the author catches is traced to the layer it came from and fixed there, in the plan, the guidance or the evaluators. Then it runs again. A fix enters the shared core only when it passes three blind tests.
every round leaves a record · plan · changelog · habits changed · review · verdicts
Authors disagree: one wants hedges in the abstract, another wants none. So your corrections never edit the core. They live in your profile, and the skill writes your way from the next paper on.
stays in your profile unless it proves general · your paper never leaves your agent
results
A NeurIPS-style preprint about peer review. Here is what the team did, and what only the author saw.
what only the author saw · paragraphs under 80 words
no reviewer or judge flagged it · traced to the plan · fixed in the constitution
SkillsBench · tasks solved · same agent
skills found through Skill Federation · agent sends a wish-list, never the paper
get it
Your agent asks skillfed.io for a research‑writing skill. Clarity & Grace comes back, team included.