Writing for Humans 1.0.1

Writing for Humans 1.0.1 moves the plugin to plugins/writing-for-humans as Quickstop becomes a Claude Code-only marketplace with every plugin at plugins/<name>/. Nothing about the style or skills changes.

• Same behaviour as 1.0.0. Apart from .claude-plugin/plugin.json, the package files are byte-identical to 1.0.0's. The manifest changes only its version and its homepage, which now points at the new path. The 1.0 quality measurements still describe this release.
• New path. The marketplace entry will point at plugins/writing-for-humans at commit 84c6a3c5c2e07e7c6f063fe6f9fb30eff6394cf4. Pins to 1.0.0 at 5581a159, path plugins/claude-code/writing-for-humans, keep working: that commit stays in history.
• Checked before release. The package installs into a fresh consumer, the style can be selected, both skills load, and every package-local reference resolves from the new install path. A fresh writing task opened with the ask, the point and the stakes. The #172 regression suite passed on this plugin tree.

Package digest 02b206d87dc28107643f98745e16484a93003931ecd61586dc1d0cc8bc1d1b6a, plugin tree 7ff44b85. The asset is the accepted source archive of 84c6a3c5. Acceptance evidence is recorded in issue #202.

When the Quickstop marketplace advertises 1.0.1, update the marketplace and the installed plugin, then run /reload-plugins or start a fresh session. Recovery uses the retained 1.0.0 release and a reviewed marketplace-pointer change; immutable releases are never overwritten.

Writing for Humans 1.0.0

Writing for Humans 1.0 makes what Claude writes easy to take in at reading speed. Every reply, report, PR description and handoff opens with the point, then what you need to do or decide, then what's at stake.

• One output style, about 570 tokens. It covers the opening, writing for a colleague who wasn't in the session, no pedantry or repetition, conclusions rather than the steps taken to reach them, and explicit asks.

• No personal voice. 1.0 drops 0.3.1's voice preset: the em-dash ban, numeral and contraction rules, the "I means the user" convention, the register split and the conventional-comments mandate. If you want those, keep them in your own custom output style.

• Skills.

  • writing-style keeps the questions and cold-reader references, and its artifacts reference is cut to what the style doesn't say.
  • writing-audit stays user-only. Its frontmatter blocks the plain forms of git push, history rewrites and gh pr merge while it runs. The README says what that can't catch.

• Measured on a frozen 9-case quality suite. Against Claude Code's built-in Concise style and 0.3.1:

  • opening cold-read: 70 of 135, against 38 and 44
  • opening stakes: 12 of 45, against 0 and 1
  • sentences led by an identifier: 7, against 46 and 43
  • median length: 309 words, against Concise's 307
  • mean cost per run: $0.0621, against 0.3.1's $0.0752

  The full gate table is in docs/acceptance/writing-for-humans-quality-baselines.md.

This is an instruction Claude follows, not a guarantee of model behaviour. The gate, the release acceptance and the retained-package evidence are recorded in issue #189.

Update the Quickstop marketplace and the installed plugin, then run /reload-plugins or start a fresh session. Recovery uses the retained 0.3.1 release and a reviewed marketplace-pointer change; immutable releases are never overwritten.

Claudit 4.2.0

Claudit 4.2.0 records why a project declined a feature, and raises a decline again only when it can show that the reason no longer holds.

• Reasons captured. Each item under "Declined by the project" now carries the reason the project gave, quoted with its file and line. A decline with no reason says "no reason given".
• Worth reconsidering, only on evidence. A declined item gets a "Worth reconsidering" note only when its stated reason matches one of a fixed set of claims and Claudit can cite evidence that the claim is no longer true. Evidence is a current Claude Code doc page showing the limitation is gone, or a repository path:line showing that the situation has changed (for example, a documented check command). Best practice is never evidence. Any other wording, or a reason with a limit or aside, gets no note. The README lists every claim and its evidence.
• Advice, not a grade hit. The note sits on the declined item, outside the ranked plan, and changes no grade. A correctness defect left in a declined feature's own configuration stays in the plan, is graded as usual and cites the decline.
• Same tree, same result. The scan decides what is declined, what is reconsidered and what is in the plan. The model writes only the explanation.

Evidence. The eval suite has four new cases: stale, valid, unsafe and no-reason declines.

• Live docs, 3 runs each: 42/42 findings, 0 false defects, every structure check passing (363/363). The new cases pass 36/36, 39/39, 39/39 and 39/39.
• Offline: 14/14.
• quickstop main, 3 fresh runs: A/A/A, "No changes worth making", Hooks declined with its reason, and no reconsider note.
• Median run: $0.13 and 38 s.

Acceptance on the retained archive is in issue #184. Source PR: #185. The marketplace pointer follows in a separate PR.

The attached source archive is the retained, accepted artifact, and the checksum identifies its exact bytes. Marketplace installation uses the accepted source commit. Model judgments are advisory. Recovery: Claudit 4.1.0 (claudit-v4.1.0, commit 70b883a) remains available.

Claudit 4.1.0

Claudit 4.1.0 gives the same result for the same tree, and follows what the project has already decided.

• Stable plan and grades. The scan, not the model, decides the grade letters, which findings become changes, their order and how many there are (up to seven). Fewer than three come with a reason, and the plan is never padded. "No changes worth making" is a valid result with Coverage A. Each grade below A names the items that would raise it.
• Consistent policy. A procedure doc that loaded team instructions already link is optional, not a change, unless the repo says to invoke it on demand. A doc that a plugin's own skill follows is product content.
• No findings on test data. Fixtures, eval cases and graders, saved reports and test directories are never read as procedures, instructions, manifests or secrets. Plugin roots are exempt.
• Documented declines. When the project documents a deliberate choice against a feature (for example "hooks were retired"), Claudit doesn't recommend it or count it against Coverage. It lists it under "Declined by the project" with the file, line and whole sentence.
• Only real calls count. A check command named in descriptive prose no longer counts for the allowlist suggestion; only commands written as calls do.
• Reproducible. The footer gives Claudit 4.1.0; audited commit <sha>. Eval and headless gate runs keep their exact command lines and every report.

Evidence. The eval suite has nine cases, including new fixture-noise, documented-decline and clean cases. Live docs: 42/42 findings, 0 false defects, every structure check passing. Offline: 14/14. Real repositories, 3 fresh clones each: quickstop efc71b7 A/A/C with the same two changes every run, and intent-to-release A/A/B. Median run $0.135 and 36 s. Acceptance on the retained archive is in issue #174. Source PR: #178. The marketplace pointer follows in a separate PR.

The attached source archive is the retained, accepted artifact, and the checksum identifies its exact bytes. Marketplace installation uses the accepted source commit. Model judgments are advisory. Recovery: Claudit 4.0.1 (claudit-v4.0.1, commit c21f6da) remains available.

Claudit 4.0.1

Claudit 4.0.1 fixes the follow-ups from 4.0.0.

• Instruction loading. Which instruction files load now comes from the scan, never from the model's memory. AGENTS.md loads when there is no CLAUDE.md, .claude/CLAUDE.md or CLAUDE.local.md in the project or above it (Claude Code 2.1.277 and later). The scan says so in one sentence, with the documented caveats. An AGENTS.md that doesn't load gets an alignment finding and the fix. .claude/AGENTS.md is discovered too.
• Plan length. Every "Top changes" section opens with "N changes are worth making". When fewer than 3 are worth making, the report gives the reason instead of padding the list.

Evidence. On the eval suite, with an AGENTS.md-only case added and F3's structural defects now script-graded: live docs 42/42, offline 14/14. No invented defects on the well-configured and AGENTS.md-only fixtures. Median run $0.136 and 50 s. Acceptance on the retained archive is in issue #168. Source PR: #169. The marketplace pointer follows in a separate PR.

The attached source archive is the retained, accepted artifact, and the checksum identifies its exact bytes. Marketplace installation uses the accepted source commit. Model judgments are advisory. Recovery: Claudit 4.0.0 (claudit-v4.0.0, commit a4342d0) remains available.

Claudit 4.0.0

Claudit 4.0.0 turns Claudit from a validator into an optimizer. One /claudit skill works out what the project is for. It grades the Claude Code setup on Correctness, Alignment and Coverage, each with a confidence. Then it recommends the 3 to 7 changes that would help most, each with the problem, the reason, the exact change and a doc citation.

• A standard-library helper (Python 3.9+) scans the setup. It reports the deterministic misconfigurations the docs name, five baseline checks and the profiles that match the project.
• The shipped catalog carries every check and recommendation. Live docs (code.claude.com only) confirm them and list pages the catalog hasn't reviewed yet. If the docs are unreachable, the report is still graded and says so in one footer line.
• Suggested allow rules are exact documented calls, never wildcards. Mutating and shell-operator calls are left out.
• It recommends /doctor, /doctor prompt-audit, /insights and /skill-doctor where they fit.
• It offers to apply selected changes in place or on a branch with a pull request. Rejected suggestions go in .claude/claudit-dismissed.json. Read-only runs write nothing.

Breaking changes. The knowledge, refresh and status skills, the research and audit agents, the docs cache and decision memory are removed. One read-only auditor agent remains. Old files under ~/.cache/claudit are no longer read, and you can delete them.

Evidence. On the #155 eval suite, the live-docs arm hit 41/42 planted findings, with one verified judge false negative, and invented no defects on the well-configured fixture. Headless audits took about 40 s and $0.13. Acceptance on the retained archive and the live-run reports are in issue #156. Source PRs: #166, plus #167 for the marketplace pointer.

The attached source archive is the retained, accepted artifact, and the checksum identifies its exact bytes. Marketplace installation uses the accepted source commit. Model judgments are advisory. Recovery: Claudit 3.2.3 (claudit-v3.2.3, commit e23d2fc) remains available.

Writing for Humans 0.3.1

Writing for Humans 0.3.1 separates readable evidence placement from the destination's formatting.

• Ordinary documents and unknown renderers default to plain Markdown sections, appendices or links. Short evidence stays nearby.
• GitHub comments and PR descriptions may use <details> for lengthy optional evidence. Other destinations may use documented native collapsible blocks when useful and supported.
• Conclusions, decisions, uncertainty and essential context stay visible. Screen-size targets no longer require collapsing or cutting necessary material.
• The output style and artifact, question and cold-reader references agree; the question example uses portable Markdown.

This is an advisory writing aid, not a guarantee of model behavior. Validation and exact-package native scenario evidence are recorded in issue #148. Claudit is unchanged.

Update the Quickstop marketplace and the installed Writing for Humans plugin, then start a fresh Claude Code session. Existing drafts are not rewritten by an update. Recovery uses the retained 0.3.0 source and a reviewed marketplace-pointer change; immutable releases are never overwritten.

Writing for Humans 0.3.0

Writing for Humans 0.3.0 clarifies native skill use and editing boundaries:

• The writing audit runs only when explicitly invoked.
• Skill references use native installed-package paths, with concrete arguments for artifacts, questions and the cold-reader reference.
• The selected output style contains less maintainer prose while retaining the substantive writing guidance.
• Installation guidance explains namespaced style selection and starting a fresh session after updates.

Native installation, reference resolution, selection, authorized editing, removal and a coding task were checked. Fixed writing trials showed unsupported factual inferences in both 0.2.0 and 0.3.0; this remains an advisory aid requiring editorial review. Reliable automatic writing-style invocation, improved factual accuracy and hot-reload effectiveness are not claimed.

The attached source archive is the retained, accepted artifact shared with the Claudit release. Marketplace installation uses the immutable accepted source commit. Recovery restores the prior 0.2.0 marketplace pointer without replacing either release's assets.

Claudit 3.2.3

Claudit now checks whether its retained documentation covers the feature being audited, even when the baseline research domains are fresh. It fetches only the missing official pages within a bounded budget, retains source provenance, and reuses fresh supplemental evidence.

Offline and no-write audits retain discovery, source reading, coverage planning and applicable auditors. Missing documentation and unobserved runtime behavior remain explicit unknowns. Research results are preserved unchanged, with failed or incomplete attempts recorded without replacing last-good evidence.

Status and knowledge retrieval expose retained supplemental pages, so later feature questions can reuse that evidence even when the baseline summary omits it. Installed no-network/no-write reuse was verified. Model workflow/reporting limitations remain documented in issue #141.

This release includes the source work from PRs #142, #144, #145 and #146. Earlier native workflow failures were retained and corrected before publication. Model judgments remain advisory; source freshness does not establish exhaustive review or runtime verification.

The attached source archive is the retained, accepted artifact. The checksum identifies its exact bytes. Marketplace installation uses the immutable accepted source commit.

Claudit 3.1.0

Claudit 3.1.0 updates configuration auditing for Claude Code 2.1.287.

• Current permission, hook, MCP, plugin and instruction discovery, with scoped decisions and explicit coverage limits.
• Fresh official-source research, independent per-domain cache freshness, and last-good preservation after failed refreshes. Persistent agent memory no longer certifies research freshness.
• Deterministic scoring that withholds an overall grade for incomplete coverage.
• Read-only boundaries, selected fixes, and isolated PR preparation that preserves unrelated work and retains failure recovery receipts.
• Six Sonnet agents and smaller, relevant evidence slices. Native projected always-on cost is about 419 tokens versus 633 in 3.0.0; main audit invocation is about 3.2k versus 7.6k. These are estimates, not measured speed gains. Model judgments remain advisory.

Requires Python 3.11+ on PATH. Native acceptance uses Claude Code 2.1.287 on macOS arm64; Windows-native behavior is not certified.

Update the Quickstop marketplace and installed Claudit plugin, then restart Claude Code. Run /claudit:status to inspect existing state and /claudit:refresh all for fresh research. Old cache and research-memory files are preserved.

The attached source archive and checksum identify the retained accepted artifact. Marketplace publication and installation verification are tracked in the linked delivery issue. Writing for Humans remains unchanged.

Implementation: #139
Delivery evidence: #138