agentleFS
ashfulcraashfulcra/fulcra-tools/skills/fulcra-content-review/SKILL.md

fulcra-content-review

Skill10 starsChanged 2 months agoMITView on GitHub

Review prose meant for humans — proposals, docs, posts — before it ships: measured AI-tell sweep, voice matching against the real author, claims discipline, and a parallel multi-lane review protocol.

The file

Show its rules and text

We split this skill into the rules it gives an agent

13 rules. About 0 are common, 12 are distinctive, and 1 is very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈12

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 1

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Necessary vs sufficient: When a test proves half a capability, say which half.

    line 39
  2. Distinctive:

    Unforgeable evidence beats self-report: Verify from the system of record (a change feed, a timestamped upload), not the actor's account of …

    line 44
  3. Distinctive:

    Numbers are measured or labeled estimates: If the colophon claims "every number was measured," make that true.

    line 55
  4. Very specific:

    Run scripts/tic-count.py (or the inline equivalent) before and after editing.

    line 72
  5. Distinctive:

    Do not sand everything flat.

    line 91
  6. Distinctive:

    Test: read a rewritten paragraph next to a real quote.

    line 102
  7. Distinctive:

    When review finds errors, there are two honest presentations: corrections left visible (credibility through shown work) or folded silently …

    line 116
  8. Distinctive:

    Ask the owner; do not default.

    line 118
  9. Distinctive:

    Reviews of a superseded draft still pay: When the draft was rewritten while reviews were in flight, triage every finding against the …

    line 134
  10. Distinctive:

    Close the loop with a reconciliation note: After applying fixes, file a map of finding → applied / resolved-by-rewrite / …

    line 139
  11. Distinctive:

    Use the author's stated style preferences and the document's purpose.

    line 149
  12. Distinctive:

    Simple and direct: Use short, connected sentences.

    line 151
  13. Distinctive:

    Make each sentence useful: Keep the argument and supporting evidence.

    line 153

≈ “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.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Fulcra Content Review

For content a human will put their name on. Code review checks that a thing works; content review checks that a thing persuades, in the owner's voice, without overclaiming. Every rule here was paid for in one real editing session — the provenance notes say where.

0. Decide point of view FIRST

Whose voice is this? Decide before any line edit — POV drives everything else. A document that says "we" (the agents), describes its owner in third person, and credits itself in the colophon is incoherent the moment the owner shares it as theirs. If the owner's voice: first person, owner's diction, agents described as tools they used. (Provenance: the Modest Proposal draft failed exactly this way.)

Pronoun-referent audit. Once POV is chosen, extract every "we/us/our" sentence programmatically and check each referent against the choice. One document, one "we" — a draft that uses "we" for the authors in one sentence and for the company being addressed in the next reads as sloppy to exactly the reader it's trying to persuade. Same failure class as everything else on this list: two names for one thing, and the text trusts the wrong one. (Provenance: v3 of the same draft shipped 8 referent slips; the owner caught them in the first few sentences.)

1. Claims discipline — before style, always

  • A single measurement is not a rate. Label it: "first live wake: 55 s", never a bare "55 s" in a stat block. One data point presented as steady-state is the most common quiet overclaim.
  • Necessary vs sufficient. When a test proves half a capability, say which half. "A scheduled run reached its tools" does not show it will act on instructions it finds elsewhere — design the second test instead of rounding up. (Provenance: the ChatGPT scheduled-task tests — the owner caught the overclaim, and the honest second test then passed cleanly.)
  • Unforgeable evidence beats self-report. Verify from the system of record (a change feed, a timestamped upload), not the actor's account of itself. Put a token in the test that exists only on the path under test — and call it read-evidence, not "unguessable"; a short hex nonce proves the file was read, it is not cryptography.
  • Preserve every leg of the evidence. Result files prove the actor did the work; only the preserved instructions prove the work wasn't smuggled in with them. Archive the prompt/charter verbatim next to the result, and state preconditions a skeptic would ask about (e.g. a one-time human pre-approval that let the run act unattended). (Provenance: second-round review caught the missing prompt artifact after the test itself had already passed.)
  • Numbers are measured or labeled estimates. If the colophon claims "every number was measured," make that true.
  • Don't infer intent from your own experience of the surface. "We didn't have to change it" does not mean "it wasn't built for this" — ease of use is usually evidence of deliberate design, not of accident. Claims about why a product is the way it is belong to its owner; check before printing one. (Provenance: a draft claimed the platform "wasn't built for agents" because the fleet rode it unmodified; the owner's correction was that the agent-facing surfaces are heavily worked precisely so models find them easy.)
  • Enumerate, don't name-match, when auditing claims. Checking a document for "the claims I remember making" has unbounded false negatives — walk every number, every stat block, every "we proved" systematically. (Same failure class as the credential sweep: four corrections because the search was scoped by a name the searcher chose.)

2. The measured tic sweep — count, then judge

Run scripts/tic-count.py (or the inline equivalent) before and after editing. Judging your own prose by taste fails; counts do not. Thresholds from a real 8,742-word draft that read as AI-written:

TellFoundTarget
Em-dashes per 100 words1.0< 0.15
"rather than"31< 8
", not X" antithesis17≤ 3 that earn it
"which is the/what/why…"13< 4
Ceremony ("worth noting/being clear/saying")14~0 — cut the ceremony, keep the point
"genuinely / precisely / exactly"16< 5
Aphorism closers on sections9 of 11≤ 2
Bold-lead definition lists5 sections≤ 2

Also hunt, unscored: triads ("cheap, standard, always available"), anaphora runs ("…is fine. …is fine. …is fine."), rhetorical questions answered immediately, every section shaped as "N things," hedging pairs ("X — and honestly, Y").

Do not sand everything flat. A few contrasts and one good image survive; the tell is density and symmetry, not existence.

3. Voice matching — calibrate on verbatim samples

  • Collect 5–10 verbatim quotes from the target author (messages, not their formal writing). Match diction, sentence length, and stance — do they ask questions straight? use contractions? think out loud?
  • Produce rewrite examples, not diagnoses. "Too formal" is useless; a before/after pair is actionable. Every reviewer returns their worst-five as rewrites.
  • Test: read a rewritten paragraph next to a real quote. If you can tell which is which by rhythm alone, keep going. (Example pair that set the bar: "We would rather be corrected than convincing." → "If any of this is wrong I'd rather hear it now.")

4. Reader-assumption check

Name the coldest intended reader, then verify the document survives them: define the platform before using it, no internal jargon without introduction, no "as discussed." If the owner says "close allies," check for facts about third-party products that are fine internally and awkward forwarded.

5. Corrections: visible or smoothed is the OWNER'S call

When review finds errors, there are two honest presentations: corrections left visible (credibility through shown work) or folded silently into a clean design. Ask the owner; do not default. Design-history sections ("how we got here, including what we got wrong") are a different thing from correction scars — they can earn their place when the journey is the argument.

6. Parallel review protocol

  • Freeze the draft at a pinned copy before fan-out; nobody reviews a moving target. One synthesis pass after all reviews land.
  • Independent reviewers, no coordination. Overlap between findings is signal (fix without debate where two agree), not wasted effort.
  • Four lanes, worst-first in each: technical (every claim vs evidence), structural (contradictions, drift between summary and body, inventory monotony), security/risk (if the content proposes systems), voice/AI-tells (lane 2–3 above).
  • The author reviews too, but grounded in counts (§2) — an author's taste pass over their own prose finds nothing.
  • Reviews of a superseded draft still pay. When the draft was rewritten while reviews were in flight, triage every finding against the current text before discarding any: some resolve by rewrite, some survive verbatim, some survive reworded. Grep the current draft for each contested passage; don't trust memory of what the rewrite fixed.
  • Close the loop with a reconciliation note. After applying fixes, file a map of finding → applied / resolved-by-rewrite / declined-with-reason, and point the next review round at it. Without it, round N+1 re-flags round N's fixes and burns its budget re-litigating.
  • Meter your own fixes. Edits made while applying review feedback reintroduce the tics — twice now, consciously avoiding them. Re-run the §2 meter after reconciliation, not just after the original edit.

7. House style for internal documents — simple and direct

Use the author's stated style preferences and the document's purpose.

  • Simple and direct. Use short, connected sentences. Rewrite passages that require repeated reading to understand.
  • Make each sentence useful. Keep the argument and supporting evidence; remove repetition. Add detail when it clarifies the reader's decision.
  • Put the ask right after the argument that motivates it, not at the end of the document. Conclusions don't need restating; one is enough.
  • Cut on sight: thesis announcements ("X is the argument"), restatements of a point a list already made, second conclusions, slogans that echo the numbered structure beside them, go-to-market speculation, and flourishes that trade accuracy for rhythm (a "three vendors" line that counted wrong shipped three review rounds before dying).
  • Keep charm that is true and the owner's. The incident-track story, the operator metaphor, the before/after tapes, "MVP first, awesome later" — reviewer taste does not outrank the owner's voice. When a reviewer flags an owner-dictated passage as slop, the reconciliation answer is "declined, operator ruling," not a compromise rewrite.
  • Sweat absolutes. "Can't be done at all" → "we couldn't do it." "Means migrating" → "can require migrating." Every absolute is a claim someone at the receiving company can falsify; write what was observed. And check the same claim's OTHER homes — the fix that lands in one section leaves its twin behind (two names for one thing, again).
  • Package for shipping. The browser-tab title, gloss for any acronym less universal than MVP, cross-refs re-checked after every structural move. Working-copy labels in shipped metadata are a real blocker class.

Scripts

  • scripts/tic-count.py <file> — strips HTML/Markdown, counts the §2 table, prints per-100-word rates. Run pre- and post-edit; the diff is the edit's receipt.

Download now (ZIP)

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
What did you use it for? e.g. Reviewing pull requests in a Rails app

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.