agentleFS
WingedGuardianWingedGuardian/skyrimvr-claude-toolkit/.agents/skills/mod-release-review/SKILL.md

mod-release-review

Skill22 starsChanged 3 days agoMITView on GitHub

Use before publishing a new version of a Skyrim mod you made — a plugin, script, mesh, asset or SKSE-DLL mod — whether a first release, an update, or a Nexus upload. Reviews only what changed since the last published version, changes nothing while reviewing, and brings every finding to you before any fix. Not for releasing this toolkit itself.

The file

We split this skill into the rules it gives an agent

10 rules. About 0 are common, 8 are distinctive, and 2 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈8

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 2

    Very specific

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

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Every rule below exists to prevent that.

    line 12
  2. Distinctive:

    Every finding goes to the author, including definite ones: No silent fixes. No finding is dropped because the agent judged it minor.

    line 22
  3. Distinctive:

    Use these two words, consistently, in everything shown to the author.

    line 49
  4. Distinctive:

    Ask, in one short message:

    line 74
  5. Distinctive:

    Read every changed or added file in full, for context.

    line 84
  6. Distinctive:

    Write the complete findings list first, with nothing applied.

    line 115
  7. Distinctive:

    When a suspected finding is in an author-tested area and the evidence is INFERRED or ASSUMED, the recommendation is to leave it, and the …

    line 129
  8. Distinctive:

    Every answer is recorded. A finding the author leaves stays left.

    line 132
  9. Very specific:

    Write the release record in a folder outside Data/ and outside the package, the same place every release (ask the author once).

    line 154
  10. Very specific:

    License notices travel with the DLL: Put the DLL's license and third-party notices beside it in SKSE/Plugins/ (<Name>-LICENSE.txt, …

    line 175

≈ “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
<!-- GENERATED by scripts/gen_agent_trees.py from agent/skills/mod-release-review/SKILL.md. Toolkit contributors: edit the source. Your own notes belong in KNOWLEDGEBASE.local.md, which updates never replace. -->

Mod release review (author-gated)

Why this review asks before it fixes. A mod's correctness is judged in the game, by its author. The agent cannot see in-game behavior. A review that runs after the author has tested in game and then "fixes" something that only looks wrong on paper can publish a broken mod that was working. Every rule below exists to prevent that.

Words used here. The author = the user publishing the mod. Players = people who install it. Author-tested = behavior the author observed in game for this version.

The rules

  1. Read-only review. While reviewing, change nothing in the mod, the game install or the package — no fixes, cleanups or "while I'm here". Exports and decompiles for reading go to a scratch folder. Reviewer subagents are told not to edit.
  2. Every finding goes to the author, including definite ones. No silent fixes. No finding is dropped because the agent judged it minor.
  3. What the author saw in game outranks the agent's reading of the files — for what they actually saw. If the code reads as X and the author observed Y, the reading is the suspect. But testing covers only the behavior observed, on the author's install, with the build they ran (Phase 1). It does not cover packaging, dependencies, other load orders or a fresh install.
  4. Nothing is fixed without the author's individual approval. Suspected findings are confirmed one at a time.
  5. A fix to an author-tested area makes that testing stale. Say so each time, and name the re-test.

When to run it

Every release, before it is published. The author chooses when:

  • before their final in-game test pass. Approved fixes then get tested by that pass. This is usually the better choice: nothing changes after the author's sign-off.
  • after their testing. Any fix approved then makes that area's testing stale (rule 5).
  • both. A full review before the final pass, then a short re-check of only what changed since.

Recommend running it before the final pass, and ask which the author wants.

Two kinds of finding

KindMeansExamples
DEFINITEprovable from the files alone; no in-game behavior involveda script fails to compile; a master the plugin needs is not a stated requirement; a file the installer names is missing; an asset that must not be published is inside the package
SUSPECTEDdepends on how the game behaves at runtimea timing, condition, property or record that looks wrong

Use these two words, consistently, in everything shown to the author.

Phase 0 — baseline: what players already have

Scope = the difference between what was last published and what will be published now. Never the whole mod.

Baseline, in order of preference:

  1. The release record of the previous version (Phase 6 writes it).
  2. The archive players actually downloaded (the Nexus file). The published bytes are the baseline, not the development files or a memory of them.
  3. A git tag, if the mod lives in a repo and the tag is known to match what was published.
  4. None of these (a first release, or nothing was kept) → the whole package is in scope.

Tell the author which baseline you used and how much changed, and get them to confirm it before reviewing. A wrong baseline quietly shrinks or inflates the review.

How to diff each file type:

  • ESP/ESM/ESL: Spriggit YAML of both sides (bash tools/spriggit-cli.sh).
  • .pex: Champollion decompile of both sides. The shipped .pex is what players' games run, not the .psc.
  • Text (.psc, INI, JSON, FOMOD XML): diff directly.
  • Binary with no text form (NIF, audio, DLL): changed = hash differs. Review that file through its tool.

Phase 1 — ask what the author tested, before judging anything

Ask, in one short message:

  • which features they tested in game for this version, and what they did not
  • which build they tested. Is it byte-identical to what will be published? Testing a development plugin, or loose files that will ship packed in a BSA, does not test the published bytes
  • anything they already know is rough

Record the answer as the author-tested set. Tag each changed area tested or untested.

Phase 2 — read the changes

Read every changed or added file in full, for context. Findings are scoped to the changed lines and what they affect. Do not hunt through unchanged files. Something wrong in unchanged files that you notice anyway is reported as PRE-EXISTING, listed separately, and does not block the release: players already have it, so holding the release does not fix it. Exception: an old defect that this release makes players REACH — through a new feature, a new default or a new line in the description — is treated as in-diff.

Phase 3 — Skyrim checklist (check the rows the diff touches)

CheckHow
Shipped .pex matches the .psc it claims to beChampollion-decompile the shipped .pex and compare with the source
Scripts compileCaprica, with the command in the skyrim-context skill
Did the author's testing exercise THIS mod's record?bash tools/skyrim-winner.sh winner <formid> on the author's load order. If another plugin wins it, their test ran the other plugin's version
Dependencies another install may lackthe plugin's masters; GetFormFromFile on another mod's forms (returns None silently when that mod is absent); SKSE plugins and frameworks at a VR-compatible build and minimum version. The author's own install has all of these, so their testing cannot catch this class
Propertiesnon-auto properties stay blank after load; added, removed or renamed properties against the plugin's VMAD
Save compatibility for players mid-playthroughnew masters; removed scripts, forms or quests; changed property sets → verdict SAFE / RISKY / DON'T
Plugin structureESL FormID range, FormID reuse, ONAM for ESM overrides
Package contentsfile list and hashes against baseline + diff; loose files vs BSA; FOMOD paths resolve to the intended Data/ layout; no source bundle inside the main package (see Packaging rules)
Content that must never be publishedthe author's private or development plugins, and assets they have no right to redistribute. Ask the author for their list once and record it
Leftover development scaffoldingdiagnostic logging, disabled test functions, unused globals
Meshescross-read with PyNifly (never PyFFI alone); no PreWEAPON/PreSHIELD nodes (VR)
SKSE DLLVR build; built from the source that was reviewed; Address Library dependency stated
Claimsversion string, changelog and the Nexus description match the files; every number re-derived
Permissions and creditsevery asset from another author: permission status and credit line

Label each finding's evidence MEASURED / READ / INFERRED / ASSUMED. Any live check in the running game (the devbench skill) happens only if the author offers a session.

Phase 4 — bring findings to the author

Write the complete findings list first, with nothing applied. Then present it:

DEFINITE findings may be shown together as one list. Each item still gets its own answer from the author. Nothing is fixed because "they were all obvious".

SUSPECTED findings: one at a time. One question per finding, never combined. Each one states:

  • what and where (file and line, or FormID), in-diff or pre-existing, evidence tier
  • whether it falls in the author-tested set, and the direct question: "Have you tested this in game? Can you vouch for it?"
  • what the agent thinks happens, that it is not certain, and the cheapest way to settle it (a log line, a decompile, a short in-game check)
  • what could break if it is changed
  • options: Leave it (I vouch for it) · Fix it · I'll test it first · Investigate more

When a suspected finding is in an author-tested area and the evidence is INFERRED or ASSUMED, the recommendation is to leave it, and the agent says so.

Every answer is recorded. A finding the author leaves stays left. It comes back only if a later release changes that code.

Phase 5 — fixes (only the ones the author approved, one by one)

  • Back up the affected files first.
  • Smallest change that does it. Rebuild (compile, deserialize). Diff against the backup and confirm that only the intended lines moved.
  • The fix is itself new, unreviewed change. Review it before it ships.
  • Verify the fix yourself first, and save the author's time for what you cannot check. In order: offline checks (compile, decompile, skyrim-winner, Spriggit diff, PyNifly), then a live check through the devbench skill if the game is running and the author allows it, within any limits the author has set on live game automation.
  • Whatever only the author can confirm in game goes on one re-test list, run as a single session at the end, not one launch per fix. Tell the author which tested area each fix touched.

Phase 6 — publish gate and the next baseline

Nothing is published until every finding has the author's recorded answer and the author gives the final go. Then:

  1. Build the package and hash its actual contents. Those hashes must match the reviewed files.
  2. Write the release record in a folder outside Data/ and outside the package, the same place every release (ask the author once). It holds:
    • SHIPPED-MANIFEST.json: every packaged file with sha256 and size, status shipped. Once the package is handed over, treat it as published: record it as the current version and the next baseline without waiting for the author to confirm the upload. The next Phase 0 checks it
    • the reviewable sources: .psc, Spriggit YAML per plugin, INI/JSON, FOMOD XML
    • REVIEW-RECORD.md: the baseline used, the author-tested set, every finding with its answer
  3. A clean review is necessary, not sufficient. The upload itself is the author's step.

Skipping step 2 makes the next release's review start from a download or a full read.

Packaging rules

  • One upload folder. Everything the author uploads or pastes (the main zip, every optional zip, any separate source zip, the page description, the summary and the changelog text) goes in the same release folder, with plain names. Give the author one path; nothing to upload lives in a build workspace or a second folder.
  • No source in the main file. The main zip holds what goes in Data/ plus the player docs at the top level. An SKSE DLL's source is its own download, never a Source/ folder inside the main zip. If its license (e.g. the GPL) requires the source to stay available, tell the author not to archive that download while the DLL ships.
  • License notices travel with the DLL. Put the DLL's license and third-party notices beside it in SKSE/Plugins/ (<Name>-LICENSE.txt, <Name>-THIRD-PARTY.txt), not at the top level, and point the README and page text at the separate source download.
  • Papyrus .psc under Scripts/Source is mod content, not "source" in this sense, and stays.

Red flags

  • About to edit anything during the review → rule 1.
  • "It's definitely broken, I'll just fix it" → present it and wait for the answer.
  • Dropping a finding because the author "probably tested that" → present it, marked author-tested.
  • Two suspected findings in one question.
  • Treating the author's testing as cover for a dependency, packaging or save-compatibility finding. Their install is not a player's install.
  • Reviewing the whole mod without telling the author there was no baseline.
  • Calling something wrong on a code read when the author has seen it work in game.

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.