invest-changelog

name	invest-changelog
description	Reads git log since the last tag and VECTOR.md value prop, groups commits by user-facing theme, filters internal noise, and writes a plain-language changelog keyed to what users can now do. Use at release time or when updating a public changelog.
argument-hint	[--since tag|commit] [--version x.y.z] [--dry-run]

Investiture Skill: Write Changelog

You are translating engineering history into user communication. git log tells you what changed in the code. VECTOR.md tells you what matters to the people using it. This skill bridges the two — reading both, filtering ruthlessly, and writing in plain language that describes outcomes, not implementation.

This skill runs at release time. It depends on a readable git history and a filled-in VECTOR.md. The richer the VECTOR.md value proposition and audience description, the more accurately the changelog can frame changes in terms users care about.

Step 1: Read Doctrine

Read the following before touching git:

1. VECTOR.md — Extract:
   • Value proposition — the core outcome the product delivers. This is the lens for deciding what is "user-facing."
   • Target audience — who reads this changelog. A changelog for developers reads differently than one for end users.
   • Design principles — informs voice and framing
   • Project stage — a discovery-stage project may not have formal releases; note this if VECTOR.md stage is "discovery"

If VECTOR.md is missing or its value prop is a template placeholder, flag it: "VECTOR.md value prop is not filled in — changelog will use plain commit grouping without audience framing." Continue.

Step 2: Determine the Commit Range

If --since [tag|commit] was passed: Use that as the start of the range. End at HEAD.

If no --since argument was passed: Find the most recent git tag.

• Run: git tag --sort=-creatordate | head -1
• If a tag exists, use it as the range start. End at HEAD.
• If no tags exist, use all commits: run git log --oneline without a range restriction.

Run git log [range]..HEAD --oneline (or git log --oneline if no range) to confirm the scope and show the commit count. Report to the terminal: "Reading [N] commits since [tag/commit]."

If the range is empty (no commits since last tag): Report: "No commits found since [tag]. Nothing to changelog." Stop.

Step 3: Read the Full Commit Log

Run: git log [range]..HEAD --format="%H %s" --no-merges

For each commit, extract:

• Hash (first 7 characters)
• Subject line (the commit message summary)
• Body if present: git show [hash] --format="%b" --no-patch

Build a working list of all commits. You will filter and group them in Steps 4 and 5.

Step 4: Filter — Drop Internal Noise

Remove commits from the working list that are not user-visible. A user cannot see these changes — they are internal to how the project is built or maintained.

Drop unconditionally:

• Dependency bumps with no behavior change: bump, upgrade, update [package], version number changes in lock files
• Build system changes: ci:, build:, chore:, changes to .github/, Makefile, Dockerfile, package.json scripts only
• Documentation changes that are not user-facing: docs:, README updates, inline comment changes, type annotation-only changes
• Formatting and linting: style:, lint:, fmt:, whitespace-only commits
• Internal refactors with no behavior change: refactor: where the subject says "no behavior change" or "internal"
• Test-only changes: test:, spec:, changes to *.test.*, *.spec.*, __tests__/

Keep these even if they look internal:

• Dependency bumps that fix a user-visible bug (e.g., "bump X to fix Y crash" — the crash fix is user-facing)
• Refactors that change performance users can observe
• Documentation changes that are part of the product (help text, UI copy, onboarding content)

Ambiguous commits: If the subject line is unclear, check the commit body for context. If still ambiguous, keep it — it's better to include something borderline than to drop a real user change.

After filtering, report: "Kept [N] of [total] commits as user-facing."

Step 5: Group by User-Facing Theme

Do not group by file, layer, or conventional commit prefix. Group by what changed for the user.

Canonical groups (use these if they fit; not all will apply to every release):

Group	What belongs here
New	Net-new capabilities the user did not have before
Improved	Existing features that work better, faster, or more reliably
Fixed	Bugs that caused incorrect behavior, crashes, or data problems
Changed	Behavior that is intentionally different in a way users will notice
Removed	Features or options that no longer exist

How to assign:

• Read each commit subject and body
• Ask: "What can a user do now that they couldn't before?" (→ New) or "What was broken that works now?" (→ Fixed)
• If a commit could fit two groups, prefer the more prominent user impact
• If a commit genuinely doesn't fit any group after filtering, drop it

Write each item in plain language. Not the commit subject — rewrite it in terms of outcome:

• fix: prevent null ref in AgentProfile.save() → "Fixed a crash when saving an agent profile with no backend selected"
• feat: add json export to agent profile → "Added JSON export for agent profiles"
• perf: cache voice list response in memory → "Voice list now loads instantly after the first fetch"

Keep each item to one sentence. Lead with the outcome, not the mechanism.

Step 6: Determine the Version Header

If --version [x.y.z] was passed: Use that version.

If no version was passed:

• Check the most recent git tag for a semver version number
• If found, suggest bumping it: "Last tag was v1.3.0. Suggest --version 1.4.0 for this release."
• If no semver tag exists, use the date: [YYYY-MM-DD]

Step 7: Write the Changelog Entry

Format:

## [version] — [date]

### New
- [item]
- [item]

### Improved
- [item]

### Fixed
- [item]

### Changed
- [item]

### Removed
- [item]

Rules:

• Omit any group with no items. Do not print empty sections.
• If only one group has items, no header needed for that group — just list the items under the version heading.
• Write for the target audience in VECTOR.md. If the audience is developers, technical specificity is fine. If the audience is end users, avoid implementation details entirely.
• Each item is one sentence. Start with a verb: "Added", "Fixed", "Removed", "Improved".
• Do not include commit hashes, PR numbers, or author names in the changelog entry itself.

Step 8: Output

Print the changelog entry to the terminal AND append it to /vector/changelog/[version].md.

Create the /vector/changelog/ directory if it does not exist.

Also check for a CHANGELOG.md at the project root:

• If it exists, prepend the new entry at the top (below the # Changelog heading if present)
• If it does not exist, create it with the new entry

Overwrite /vector/changelog/[version].md if it already exists for this version — the current content is what matters.

If --dry-run was passed: Print the entry to the terminal only. Do not write any files.

After writing, output:

Changelog entry written to /vector/changelog/[version].md
[If CHANGELOG.md was updated:] CHANGELOG.md updated.

[N] commits in range. [N] kept as user-facing. [N] dropped as internal.

Arguments

• No arguments: Since last git tag, version from last tag or date
• --since [tag|commit]: Explicit range start
• --version [x.y.z]: Set the version header for this entry
• --dry-run: Preview the entry without writing files

Output

• /vector/changelog/[version].md
• CHANGELOG.md (prepended if exists, created if not)

When to Run

• At release time, before publishing release notes or updating a public changelog
• After a sprint, to communicate what shipped to stakeholders
• Before tagging a release — use --dry-run to preview what the release notes will say

Principles

• Value prop is the filter. If VECTOR.md says the product's value is "instant, portable agent voices," a CSS refactor is not in the changelog. A new voice backend is.
• Outcomes, not mechanisms. "Fixed null reference in AgentProfile" is a mechanism. "Fixed a crash when saving an agent profile with no backend selected" is an outcome. Rewrite for the latter.
• Omit empty groups. A changelog with six headers and one item each is noise. A changelog with two headers and five meaningful items is signal.
• git log is evidence, not the changelog. Commit messages are for engineers. The changelog is for users. They are different documents with different audiences. Translate, don't copy.
• Drop noise ruthlessly. A changelog that includes every internal refactor trains users to stop reading it. The value of a changelog is in what it omits as much as what it includes.