Four Claude Code plugins for making things, and checking them.
 |
 |
 |
 |
| story-forge writes the book |
design-forge designs the site |
deck-forge builds the deck |
proof-forge checks the export |
Same guy. He makes the thing, then he checks the thing.
| design-forge | Visual work — an adversarial builder/critic loop, a measurement harness, a licence-cleared asset library, a scroll-film site builder |
| story-forge | Long-form fiction — braindump to dossier, outlining, drafting, a three-pass line edit, a continuity audit |
| deck-forge | Presentations — a component library that never builds a native chart, and a harness that measures legibility at distance and contrast against the real backdrop |
| proof-forge | Exports — compare two renders and mark exactly where they differ, on video, stills or audio |
They share one idea: judgment stays with the model, and anything expressed as a number gets measured instead of eyeballed.
That is not a style preference. Vision models score 58.57% on trivial geometric tasks and 56.84% on counting line intersections in an image. Open-ended AI design audit runs an 80.1% false-positive rate, with 8.9% of it actively harmful advice. Ask a model whether a page is any good and you get something useful. Ask it to count accents or verify a hex and you get confident noise.
So the work is split. A harness computes what is computable. The model is spent on taste, structure, and whether the thing is actually any good.
Why that split is worth the trouble: a 2026 survey of 1,000 workers found 46% say fixing AI output takes as long as doing the work manually, and 57% say the time saving disappears once corrections are needed. METR's RCT is sharper still — experienced developers ran 19% slower while believing they were 20% faster. AI collapses generation time and inflates verification time. The harness decides which one wins.
Sources and verification tiers for every figure: design-forge/references/mechanisms.md
claude plugin marketplace add treynortetik-creator/forge
claude plugin install design-forge@forge # any of them,
claude plugin install story-forge@forge # or
claude plugin install deck-forge@forge # or
claude plugin install proof-forge@forge # all fourOr clone and install locally:
git clone https://github.com/treynortetik-creator/forge && cd forge
claude plugin marketplace add ./
claude plugin install design-forge@forgeRestart the session or /clear afterwards so the skills load.
design-forge has an optional toolchain (ImageMagick, Inkscape, Graphviz, ffmpeg, DuckDB) — see
design-forge/scripts/install.sh --tools and design-forge/scripts/doctor.sh, which reports what is
missing and what each gap actually costs you. Neither plugin requires it; both run on python3 and
a browser.
Eight skills. The one that earns its keep is design-audit — a read-only browser harness that
computes what a screenshot cannot show.
Three pages once went through four rounds of the critic loop and were declared finished. The harness then found 23 real defects in under two minutes: a colour that was simply the wrong hex, seventeen WCAG failures, a type-ladder violation, tap targets under the legal minimum. Then a red team found bugs in the harness, and the fixed version found more still on the same "clean" pages.
Four things a screenshot structurally cannot do:
- Verify an exact value. Two lavenders forty units apart look identical.
- Falsify a per-viewport rule. A screenshot is one viewport, and you chose which one.
- Compare two numbers far apart in the page. A 655px hero and a 688px story column never share a frame.
- Compute a ratio. Nobody eyeballs 4.21 against 4.50.
What it does not see, stated plainly: one page, one state, one width. No hover or focus states (SC 2.4.7 Focus Visible is Level AA and is unchecked), no dark mode, no error or empty states, no mobile, no RTL, no print, nothing about the accessibility tree. Two AA criteria implemented out of roughly fifty-five. A green report means conformant on the axes measured — not accessible, and not good.
photo-to-3d-loop applies the same thesis to geometry: model a real object in Blender from
reference photographs, render fixed views, measure the silhouettes, and send a fresh critic that
never saw the build. Twenty rounds reached 91.9% mean IoU, and about half of them went into finding
out that an instrument was lying rather than that the model was wrong. Its hardest-won rule is that
a person who has physically handled the object outranks every critic and every metric — the
defect that closed the project was flat flanks on a part that curves, invisible to four rounds of
critics and to the metric, because a flat flank and a convex one cast the identical silhouette from
every view the rig could render.
Seventeen skills and four chaining commands, from a braindump through to a manuscript that has been line-edited, continuity-audited, and stripped of invisible provenance characters.
The editing half is shared with design-forge: de-sloppifier (a three-pass line edit),
voice (extract a style fingerprint from real samples so the prose is anchored to a person
instead of the model's defaults), and clean-export.
de-sloppifier leads with present participial clauses, not a word list — PNAS measured
instruction-tuned models using them at 5.3× the human rate. The durable machine signature is
grammatical, not lexical; word lists date fast and survive light editing.
It also deliberately does not use the word burstiness. That term has three incompatible meanings, the popular one is a mutation of a vendor coinage, and the only peer-reviewed test of it found it unreliable. What replaced it is honest and still useful: sentence-length CV, where human non-fiction runs ~78% against ~50% for machine prose at nearly identical mean sentence length. The signal was never length. It was dispersion.
A native PowerPoint chart flattens into a static image on Google Slides import, and there is no
workaround inside an uploaded .pptx — Slides only keeps a chart live when it is linked to a Sheet,
and that link cannot exist in a file you upload. The recipient cannot fix a typo in an axis label.
The usual response is a style guide telling people to avoid charts. Style guides lose. So
deck_build.py has no add_chart, and every visual is built from rectangles,
ovals, text boxes and tables — the primitives that survive the import editable.
Slide.s and Deck.prs are public, so a native chart is one
attribute hop away. An earlier draft of this README claimed compliance was "structural"; a review
disproved it in one line. What the round trip guarantees is narrower and still useful — the library
never creates a chart, and the audit catches one if you reach around it. The builder also enforces the legibility floor with an exception rather than
a warning, on the theory that a warning you can scroll past is how 8pt footnotes ship.
deck_audit.py measures what a person looking at a slide cannot:
| legibility | minimum point size from the ANSI/INFOCOMM V202.01 element-height model |
| contrast | WCAG 1.4.3 against the actual backdrop, resolved by z-order |
| native charts | the thing that will silently become a picture |
| off-slide geometry | including shapes nested inside groups |
Two results worth stating plainly. Points are not a physical size — a point is a document unit, and treating it as an inch gives a 185pt floor for a 30ft room. And WCAG large text is 18pt or 14pt bold — the familiar 24px/18.66px figures are those same sizes in CSS pixels, so applying them to points invents failures between 18 and 24pt.
/200 acuity factor is straight out of ANSI/INFOCOMM
V202.01 §4.3.1. The 4/6/8 rule is not — it appears zero times in the standard, and AVIXA's own
training deck titles it "The Old Way of Doing Things" with "Origins are unclear." The arithmetic
is sound; the provenance is folklore. An earlier draft of this README claimed the opposite, and also
claimed the resulting number explains where the folk "24pt minimum" came from. Nothing supports that,
and the claim is withdrawn.
The harness shipped four times reporting green on the defect it exists to catch, including a version that returned "0 runs, all PASS, exit 0" on a deck built exactly the way the plugin says to build one. Each of those is now a regression test.
The single strongest request in a dig through practitioner forums, verbatim:
"I need to compare two exports and make sure they are exact visual replicas... the only way I know how is to drop a video file into the top layer and compare it one clip at a time. This is incredibly tedious and isn't even foolproof. Sometimes the difference is just one little corner of the frame for one second in a two-hour sequence."
And the spec for the fix, from the same person: "even if there was a tool that just did the difference-mode scanning for you and added markers where the differences appeared, that would be a big step." The commercial equivalent starts in the mid five figures.
The hard part is the threshold, not the diff. Two exports of the same timeline are never bit-identical — a different encoder, bitrate or colour pipeline moves every pixel slightly. An absolute threshold therefore flags either the whole file or none of it, depending on codec, and a tool that cries wolf on a routine re-encode is one people stop running.
So the threshold is derived from the pair being compared: take the median per-frame PSNR as that pair's own encoding-noise floor and flag frames that fall well below it. Verified in both directions — CRF 18 against CRF 32, and against an entirely different codec, produce zero markers; a 70×70 box for one second in a 640×360 frame produces exactly one, in the right second, with an EDL you can import.
Stills get a bounding box and a per-tile heat map. Audio gets a phase-invert null test, which is exact rather than statistical: identical audio cancels to digital silence, and a 0.1% gain change surfaces at −78 dBFS.
31 tests, fixtures generated with ffmpeg at run time.
Both plugins stand on other people's work and say so.
- design-loop is adapted from The Design Loop, itself a variation on the Gauntlet Loop originated by Matt Shumer. The method is his.
- story-forge's craft notes summarise publicly posted material, largely Jason Hamilton / The Nerdy Novelist, plus Wulf Moon's framework and Browne & King's Self-Editing for Fiction Writers. Each note names its source in frontmatter. Full detail: story-forge/NOTICES.md.
- 🔴 The de-sloppifier's AI-vocabulary lists overlap Wikipedia's Signs of AI writing, which is CC BY-SA — attribution and share-alike required. Credited in both plugins' notices. If you reuse those lists, carry the attribution.
- Two notes were deliberately removed before this went public because they are third-party creative assets rather than summaries of a method — a verbatim prompt and a curated vocabulary list. Nothing is broken by their absence; see story-forge/references/writing/README.md.
- Anthropic's
frontend-designskill is quoted in design-forge's house-style notes, marked inline. - The asset library ships 61 files under MIT, ISC, OFL-1.1, CC0 and public-domain terms, each with its licence and where that licence was verified recorded per file. See design-forge/THIRD-PARTY-NOTICES.md.
If you are one of the creators above and want something changed or removed, open an issue and it goes the same day.
MIT for the original work in this repository. The bundled assets under
design-forge/skills/art-department/library/ remain under their own licences, which carry
redistribution obligations of their own — those are enumerated in
design-forge/THIRD-PARTY-NOTICES.md.