Proposal: maintainer-reviewed Zero Slop portable writing skill bundle

[prompeteer]

Proposal

Please consider a maintainer-reviewed import of Zero Slop, an MIT-licensed writing skill maintained at manavmishra/ZeroSlop. I am using the Ideas route required by the current contribution guide for bundles containing scripts, before attempting a fork PR. This is a request for review, not a claim of inclusion in AAS.

Affiliation: this proposal comes from the Zero Slop author's authorized agent workflow.

Task and gap

Zero Slop guides an existing assistant through detecting stock AI phrasing, editing a draft, checking the result against the source, and optionally learning a person's expressly approved editing preferences. Its local helpers check names, numbers, quotations and links, alongside heuristic writing patterns. The score measures surface patterns; it does not identify an author or prove writing quality.

I checked the current source catalog and nearby unslop, unslop-file, unslop-review and copy-editing skills. unslop documents a separate CLI for post-processing; copy-editing supplies editorial passes. The proposed addition supplies a bundled offline checking workflow with source-detail checks, a no-Python fallback, and private preference handling. These areas overlap, so maintainer judgment on whether the added capability warrants a separate catalog entry would be useful.

Example task: “Edit this announcement without changing the claim: ‘We're thrilled to announce that our team has leveraged cutting-edge machine learning to deliver a seamless onboarding experience, reducing setup time by 40%.’” The acceptance condition is to remove stock phrasing while retaining the 40% claim and avoiding invented evidence. This is an illustrative test input, not a newly measured agent result.

Proposed source and complete bundle

Pin the initial review to upstream commit 01e99404ef02062fc6df4620cf7ac89014769c77, whose skill metadata is version 2.12.32.

The proposed portable runtime is the complete skills/zero-slop/ bundle, plus its retained upstream MIT LICENSE. It contains SKILL.md, references/, data/, and these eleven Python helpers:

• calibrate.py
• learn.py
• predictability.py
• reader_review.py
• register.py
• rerank.py
• rescue.py (upstream executable mode)
• safeio.py
• slopscore.py
• text_hygiene.py
• version_check.py

I would retain their real names and disclose them for protected review. I would not rename executable content as data or submit a partial copy whose relative references are broken. Packaging, website, benchmarks, maintainer-only research utilities, hosted service code, and host-specific MCP adapters are outside this proposed portable import. A maintainer-sponsored PR would add the required AAS provenance fields and retained license notice, with any adaptations listed against the pinned source.

Operational boundaries

• Dependencies: the Python helpers use the standard library. If Python 3 is unavailable, the skill has a documented manual rubric fallback. An existing AI assistant performs the editorial inference.
• Data sent: the local checks do not transmit drafts. The optional version_check.py makes one metadata-only GitHub release request, can be disabled with ZS_NO_UPDATE_CHECK=1, and is separate from draft processing. The host assistant's own privacy settings still apply.
• Writes: scoring can keep an optional local run counter under $ZERO_SLOP_HOME; preference/voice changes require the relevant explicit user instruction. Other helpers can write an explicit selected output, and maintainer calibration/merge paths require separate review. No private learned state is part of the proposed import.
• Hosted alternatives: the upstream npm deslop command and hosted MCP/REST service send a draft remotely and have bounded shared capacity. They are opt-in alternatives, not prerequisites for the portable local workflow, and the skill forbids invoking them during an offline run without the user's request.
• Costs/accounts: this proposed local bundle needs no Zero Slop account or API key. Use of the user's existing AI assistant follows that assistant's terms and costs; the proposal performs no paid inference.

The detailed published boundaries are in upstream SECURITY.md.

Evidence and requested next step

For this proposal I inspected the pinned public skill, its bundle inventory, packaging exclusions, security contract, current AAS contributor rules and overlapping skills. Searches found no existing Zero Slop PR, issue or Discussion. I have not run AAS's complete contribution preflight or an end-to-end agent edit for an AAS import, and source inspection is not a passing semantic review.

Could a maintainer confirm whether this complete bundle is suitable for a sponsored protected owner PR? If so, the follow-up should include the exact source diff, complete current preflight, practical host test evidence and manual review against the final candidate head. No generated catalog, plugin mirror or workflow changes are proposed in a normal fork PR.

[sickn33]

Thanks for the detailed proposal. We reviewed the proposed portable bundle at 01e99404ef02062fc6df4620cf7ac89014769c77 and think it is worth preparing an AAS integration candidate. The local source-detail checks and rewrite comparison add something useful beyond the existing prose-editing skills; 82 focused tests passed in our local evaluation.

Please prepare one complete zero-slop skill bundle with the required references, data, Python helpers, retained MIT notice and AAS provenance metadata. Keep the hosted service, MCP adapters, website and distribution tooling outside this import.

For the AAS adaptation, please address these points:

• Use paths relative to the installed skill and keep private preferences/state outside the distributed bundle, with learning enabled only by an explicit user request.
• Disable the upstream update check by default for this packaged copy, or adapt it to the AAS distribution. It should not tell AAS users to overwrite their installation with npx skills add manavmishra/ZeroSlop --global.
• Preserve the distinction between heuristic checks and semantic verification. A preserved set of names and numbers is not proof that a rewrite preserves the claim; the contextual comparison must remain explicit.
• Include the exact upstream revision, adaptation diff, relevant test results and practical host-test evidence, with any untested behavior stated clearly. If you use the newer upstream workflow/reference split, identify that revision and its changes rather than silently changing the proposed base.

Because this bundle contains executable helpers, please share the prepared branch and integration notes here first so we can handle the candidate through the protected maintainer review path described in CONTRIBUTING.md. We will review the complete final bundle and checks before deciding on inclusion.

[manavmishra]

Thanks for the review. The candidate branch is ready for the protected maintainer path: https://github.com/manavmishra/agentic-awesome-skills/tree/add-zero-slop

I have not opened a PR, because the bundle includes scripts and you asked to see the branch first.

Revision. Upstream manavmishra/ZeroSlop at f9e236bb55ef9c4fbe9d258e9b1f1a8dc258e8c2 (v2.12.33), not the proposed 01e99404 (v2.12.32). Within skills/zero-slop/, the only difference between them is the workflow split: SKILL.md went from 65,548 to 14,202 characters, and its detailed workflow moved unchanged into references/detailed-workflow.md. Commit 1 on the branch is that tree plus the upstream MIT LICENSE, byte for byte.

Adaptations (commits 2 and 3, on the branch):

• Frontmatter: added risk: safe, source, source_repo, source_type: community, date_added and license_source. Use self instead if you prefer.
• scripts/version_check.py makes no request unless ZS_UPDATE_CHECK=1 is set, and the npx skills add ... --global hint is gone. Step 0 of the workflow no longer runs a check on its own.
• To pass npm run validate, I shortened the description from 474 to 239 characters and added a ## When to Use section.
• Private state stays under $ZERO_SLOP_HOME, outside the bundle, and learning still needs the user's separate opt-in. Both are unchanged from upstream. The heuristic checks and the contextual comparison are still described as separate steps.

Results.

• npm run validate passes for all 2,903 skills, and validate:references passes. security:docs finished with no findings.
• pr:preflight --contributor fails with 92 new_unapproved_path findings out of 105 changed files: 11 scripts, 3 data/ files and 78 files under data/corpus/. SKILL.md, references/ and LICENSE raise none. The corpus files are test fixtures that my proposal did not call out. Say if you would rather I drop them.
• The 11 helpers compile. slopscore.py flagged 4 phrases on the example sentence from the proposal, and version_check.py stays offline by default.

Not tested. The opt-in network path of version_check.py, an end-to-end edit in a real agent host, and upstream's own test suite against the adapted copy. No PR description exists yet, so the evidence check has not run.

One more note: slopscore.py can print a one-time GitHub-star request in an interactive terminal. I left it as upstream has it. I can remove it for the AAS copy.

[manavmishra]

Follow-up with the checks I listed as not yet run. Branch unchanged: https://github.com/manavmishra/agentic-awesome-skills/tree/add-zero-slop

Host test. I ran the adapted skill in a Claude session on the example from the proposal. The edit was: "Our team used machine learning to build the onboarding experience, reducing setup time by 40%." The writing score went from 99.3 (4 flagged phrases) to 9.8 (none). slopscore.py --fidelity reported the facts preserved and nothing added, with the one figure kept. I also compared the edit with the source myself: the machine-learning use, the onboarding subject and the 40% setup-time reduction survive. The "We're thrilled to announce" opener, "cutting-edge" and "seamless" are gone. This is one sentence in one host, not a benchmark.

Update check. With ZS_UPDATE_CHECK=1 and a working certificate bundle, version_check.py returns 2.12.33 with checked: true. Without certificates it fails open, as designed. ZS_NO_UPDATE_CHECK=1 still overrides it.

Upstream test suite. I ran its 588 tests on the 2.12.33 archive with the two adapted files overlaid. Five failures also occur on the unmodified archive, because an archive export has no git history. Two more fail because I edited the workflow file without rebuilding upstream's dist/ bundle; the AAS bundle has no dist/. Everything else passes.

Still open: a passing pr:preflight and the PR-description check, both of which need the sponsored PR.

[manavmishra]

Correction to my last update. I reran the upstream suite on a full-history clone of the 2.12.33 commit instead of a git archive export. Unmodified, all 588 tests pass (1 skipped). With the adapted version_check.py and detailed-workflow.md overlaid and dist/ rebuilt, all 588 pass (1 skipped).

So the seven failures I reported earlier came from my test setup: the export had no git history, and I had not rebuilt dist/ after editing the workflow file. They were not defects in the adaptation. I had given the git-history cause as likely but had not verified it; now it is.

The adaptation is not applied to upstream, so this shows only that it breaks none of upstream's tests. The earlier items still open are a passing pr:preflight and the PR-description check, which need a sponsored PR.