marginal-contribution
Skill2 starsChanged 5 months agoMITView on GitHub
Score and improve a piece of writing for MARGINAL CONTRIBUTION — what it adds over prior work and the live thread, not polish. Use before declaring any substantive writing "done" (whitepaper, paper, post, reply, design doc, README), or when the user says "score this for contribution", "is this worth sending", "cut the redundancy", "run the writing loop / marginal-contribution loop", "mc this". The writing-side twin of /critical-qa (which is the code side). Operationalizes primitive_marginal-contribution-evolution-loop. Do NOT use for code (use /critical-qa, /code-review, /verify, /simplify) or for grammar/tone polish.
In this repository7 other files it gives its agents.
- classifySkill
- critical-qaSkill
- house-styleSkill
- literature-loopSkill
- md-to-pdfSkill
- ship-webSkill
- trp2Skill
The file
We split this skill into the rules it gives an agent
3 rules. About 0 are common, 3 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈3
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 36
Run these as defect-hunts, with the prior work in hand.
- Distinctive: line 41
If the plain concept is already stated and the name adds zero concept, CUT the name.
- Distinctive: line 53
When choosing between variants, ask which contributes more, not for a score.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
marginal-contribution — score writing by what it ADDS, not how polished it is
Operationalizes primitive_marginal-contribution-evolution-loop (writing instance) and
Desktop/marginal-contribution-writing.md. This is the writing-side twin of the code loop already
wired to /critical-qa /code-review /verify /simplify. The thesis: the value of writing is
its marginal contribution over everything already said. A flawless paragraph that restates known
material adds nothing. Optimize for the gap, never over-claim.
Inputs to gather FIRST (the loop is theater without these)
The judge must see the reference frame, not just the draft:
- The prior work the piece enters — the existing thread/papers/corpus/codebase on the topic. If a
path or thread is given via
--against, read it. Otherwise ask the user for it or name what you used. Do not score contribution against nothing. - What the reader already knows. Audience and the live conversation they are in.
- The draft itself (the real artifact, never a summary).
The two axes (BOTH required, gate A before climbing B)
- AXIS-A — contribution to the CONVERSATION. Reference = all prior work the audience already holds. Question: is this worth existing / sending AT ALL? Gate this FIRST. If A ≈ 0 (the draft restates what's already out there), STOP and say so — no amount of B-polish redeems polished redundancy. This is the most common failure of competent writing.
- AXIS-B — contribution over the LAST ITERATION. Reference = the previous draft. Question: is this revision improving, and has it converged? Climb B only after A passes. Converge when a round cannot beat the one before it (B-Δ → 0).
The judge rubric (score by BREAKING, not by rating 1–10)
Run these as defect-hunts, with the prior work in hand. Rating rationalizes; hunting for the flaw catches.
- REDUNDANCY (axis-A). Given the full prior exchange, list every draft claim already present in it. High overlap ⇒ cut it or compress to a single line; put the document's weight on the gap.
- CREDENTIAL-DROP. Flag every named theorem/framework/buzzword (Shapley, VCG, Myerson, "dogfood"…). If the plain concept is already stated and the name adds zero concept, CUT the name.
- OVER-CLAIM. Flag any open question stated as settled, or an idea dressed as a proof. The honest move is to label a conjecture a conjecture. (This is what landed the whitepaper's "load-bearing open problem rather than a solved one" line — keep that discipline.)
- CALIBRATE (verify-the-verifier — the keystone). Plant one known instance of each failure above into a probe draft; confirm the judge flags it. A judge that misses planted defects is theater, and its green light is worse than no judge.
Judge hygiene
- Maker ≠ checker. Score with an independent perspective, not the author's. Decorrelate frames; N copies of one frame shallow-agree and are wrong together.
- Pairwise > absolute. Judges are reliable at "A vs B: which adds more"; unreliable at "rate 1–10". When choosing between variants, ask which contributes more, not for a score.
- Correctness is a HARD GATE, above usefulness/novelty/Δ. A false claim fails regardless of polish.
The loop
- Gate AXIS-A. If it fails, report that and stop (the most valuable output is often "don't send this").
- Generate a few variants concentrating on the uncovered gap (cut restatement, compress known parts to single lines).
- Score each by the rubric above with independent judgment.
- Synthesize the winner. Repeat until B-Δ → 0.
- Output: (a) the cut list — what was redundant/credential-drop/over-claim, (b) the gap the
piece should concentrate on, (c) if
--rewrite, the synthesized draft. Always show the cut list, not just the rewrite — the cuts are the contribution.
Scope
Writing only. For code use the oracle-scored twin (/critical-qa, /code-review, /verify,
/simplify). This is not grammar or tone polish; it is contribution measurement.
North star
Distill what a competent human reviewer would catch into mechanical checks the loop runs without them (pairwise ground-truth, watched-repo merge/reject signal). A loop is only as good as its verifier; the goal is a verifier that inherits a human's catch-rate so it catches at scale, author out of seat.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

