Claude Code config: smaller start-up context for this repo #110

Merged

claude-bot merged 4 commits from chore/claude-config into main 2026-10-29 08:34:35 +00:00

Conversation 4 Commits 4 Files changed 3 +60

claude-bot commented 2026-10-29 08:06:56 +00:00

Collaborator

• Copy link
• Copy Markdown

Not merge-ready on its own: the AI-behavior gate (an eval suite) is not run because the repo has none. Whether to merge without it is the owner's call.

What

Claude Code sessions in this repo start with less context. Project settings turn off plugins and skills that this repo doesn't use (superpowers, browser, docs, artifacts, dataviz and others) and deny the Artifact tool. A utbible-worker agent with a short tool list runs the utbible-feature workflow stages. CLAUDE.md gains a "Keeping context small" section, which says how to turn superpowers back on for a bigger feature (claude --settings …).

Blast radius

Tier: leaf
Trunk paths touched: none
Testing tier: AI behavior (also: ship on sight, for the CLAUDE.md docs)
Gate / rollback: revert this merge. Nothing in the app, its build or CI reads these files.

Proof

• settings.json is valid JSON and the agent frontmatter parses (name, description, tools): python3 -I checks pass.
• The settings take effect: a fresh claude -p --model haiku session in the worktree listed no superpowers, artifact, dataviz, docs or pdf skills, listed utbible-worker among its agent types, and answered "no" when asked whether it had an Artifact tool.
• The per-session opt-in works: the same session started with --settings '{"enabledPlugins":{"superpowers@claude-plugins-official":true}}' listed all 14 superpowers:* skills.
• CI: run on e6f87a8 (see checks).
• not run: AI-behavior eval suite, because the repo has none. The first utbible-feature run on the worker agent is the real test.

Review

Reviewer: fresh-context code-review at medium (review 300, on 08b35cc): 3 findings, all low. 2 resolved, 1 left open:

• fixed (f83ef13): CLAUDE.md now includes ToolSearch in the worker's tool list.
• resolved by design (owner comment 2265, reply 2267): with superpowers off, babysitting-a-pr can't load its superpowers skills. Small PRs go without them; a bigger feature turns superpowers on per session (e6f87a8).
• open: with Artifact denied, using-a-design-system can't read an Artifact-hosted design system. This matters for #95 (design and rebrand). Remove the deny for that work.

Review-Confidence: not audited (leaf)

Confidence by part

• settings.json: high. Checked in a live session, with and without the opt-in.
• utbible-worker agent: medium. The tool list matches the workflow's stages according to the review, but no workflow has run on it yet.
• CLAUDE.md text: high.

🤖 Generated with Claude Code

> **Not merge-ready on its own:** the AI-behavior gate (an eval suite) is not run because the repo has none. Whether to merge without it is the owner's call. ## What Claude Code sessions in this repo start with less context. Project settings turn off plugins and skills that this repo doesn't use (superpowers, browser, docs, artifacts, dataviz and others) and deny the Artifact tool. A `utbible-worker` agent with a short tool list runs the `utbible-feature` workflow stages. CLAUDE.md gains a "Keeping context small" section, which says how to turn superpowers back on for a bigger feature (`claude --settings …`). ## Blast radius **Tier:** leaf **Trunk paths touched:** none **Testing tier:** AI behavior (also: ship on sight, for the CLAUDE.md docs) **Gate / rollback:** revert this merge. Nothing in the app, its build or CI reads these files. ## Proof - settings.json is valid JSON and the agent frontmatter parses (name, description, tools): `python3 -I` checks pass. - The settings take effect: a fresh `claude -p --model haiku` session in the worktree listed no superpowers, artifact, dataviz, docs or pdf skills, listed `utbible-worker` among its agent types, and answered "no" when asked whether it had an Artifact tool. - The per-session opt-in works: the same session started with `--settings '{"enabledPlugins":{"superpowers@claude-plugins-official":true}}'` listed all 14 `superpowers:*` skills. - CI: run on e6f87a8 (see checks). - not run: AI-behavior eval suite, because the repo has none. The first `utbible-feature` run on the worker agent is the real test. ## Review **Reviewer:** fresh-context `code-review` at medium (review 300, on 08b35cc): 3 findings, all low. 2 resolved, 1 left open: - fixed (f83ef13): CLAUDE.md now includes ToolSearch in the worker's tool list. - resolved by design (owner comment 2265, reply 2267): with superpowers off, `babysitting-a-pr` can't load its superpowers skills. Small PRs go without them; a bigger feature turns superpowers on per session (e6f87a8). - open: with Artifact denied, `using-a-design-system` can't read an Artifact-hosted design system. This matters for #95 (design and rebrand). Remove the deny for that work. **Review-Confidence:** not audited (leaf) ## Confidence by part - settings.json: high. Checked in a live session, with and without the opt-in. - utbible-worker agent: medium. The tool list matches the workflow's stages according to the review, but no workflow has run on it yet. - CLAUDE.md text: high. 🤖 Generated with [Claude Code](https://claude.com/claude-code)

claude-bot added 2 commits 2026-10-29 08:06:56 +00:00

Claude Code config: smaller start-up context for this repo ... 2d6116b8ef

- .claude/settings.json: deny Artifact, turn off plugins and skills this
  repo never uses (superpowers, typescript-lsp, playwright, office/browser
  and artifact skills). Measured: main first context 28.0k -> 24.9k, plus
  ~9.4k from the Artifact schema in main and every agent.
- .claude/agents/utbible-worker.md: agent for workflow stages with a
  short tool list (first context 20.2k -> 12.3k).
- CLAUDE.md: "Keeping context small" (gate output, screenshots, tables).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Merge remote-tracking branch 'origin/main' into ship/claude-config 08b35cc50b

claude-review reviewed 2026-10-29 08:07:50 +00:00

claude-review left a comment

• Copy link
• Copy Markdown

Review level: medium (leaf). Commit: 08b35cc50b. Range: origin/main...HEAD.

Files reviewed: .claude/agents/utbible-worker.md, .claude/settings.json, CLAUDE.md, ~/.claude/workflows/utbible-feature.js (consumer), agent-workflow skills (shipping-a-pr, babysitting-a-pr, auditing-before-launch, using-a-design-system).

No high or medium findings. The workflow passes agentType 'utbible-worker'. Its stages need only Bash/Read/Edit/Write/Skill/ToolSearch and the two MCP servers, which the tool list provides. Review and audit run in one pass with no subagent spawn, so the missing Agent tool does not break them. Settings keys and values are valid.

Low:

• .claude/settings.json:9 (superpowers plugin off) · agent-workflow babysitting-a-pr/SKILL.md:36,38 tells the agent to use superpowers:systematic-debugging and superpowers:receiving-code-review. In this repo those skills are off, so a babysit run on red CI or review comments cannot load them and falls back to ad hoc handling. The same goes for the workflow's "TDD where it fits" with superpowers:test-driven-development.
• .claude/settings.json:4 (Artifact denied) · using-a-design-system/SKILL.md:45 reads design systems with the Artifact tool. A UI or QML task that triggers that skill in this repo cannot read the design system and has to stop and say so.
• CLAUDE.md:119 · the text says the worker loads only Bash, Read, Edit, Write, Skill and the two MCP servers. The agent file also lists ToolSearch, which the MCP stages need. This is harmless but inexact.

Review level: medium (leaf). Commit: 08b35cc50b8287216ed5853264203985131b4f3f. Range: origin/main...HEAD. Files reviewed: .claude/agents/utbible-worker.md, .claude/settings.json, CLAUDE.md, ~/.claude/workflows/utbible-feature.js (consumer), agent-workflow skills (shipping-a-pr, babysitting-a-pr, auditing-before-launch, using-a-design-system). No high or medium findings. The workflow passes agentType 'utbible-worker'. Its stages need only Bash/Read/Edit/Write/Skill/ToolSearch and the two MCP servers, which the tool list provides. Review and audit run in one pass with no subagent spawn, so the missing Agent tool does not break them. Settings keys and values are valid. Low: - .claude/settings.json:9 (superpowers plugin off) · agent-workflow babysitting-a-pr/SKILL.md:36,38 tells the agent to use superpowers:systematic-debugging and superpowers:receiving-code-review. In this repo those skills are off, so a babysit run on red CI or review comments cannot load them and falls back to ad hoc handling. The same goes for the workflow's "TDD where it fits" with superpowers:test-driven-development. - .claude/settings.json:4 (Artifact denied) · using-a-design-system/SKILL.md:45 reads design systems with the Artifact tool. A UI or QML task that triggers that skill in this repo cannot read the design system and has to stop and say so. - CLAUDE.md:119 · the text says the worker loads only Bash, Read, Edit, Write, Skill and the two MCP servers. The agent file also lists ToolSearch, which the MCP stages need. This is harmless but inexact.

schaefera added 1 commit 2026-10-29 08:08:14 +00:00

CLAUDE.md: the worker also loads ToolSearch (review 300) ... f83ef136ca

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

schaefera commented 2026-10-29 08:12:41 +00:00

Owner

• Copy link
• Copy Markdown

I wonder if superpowers should be conditionally used for bigger features, as it's rather a waste for small bugfixes or config toggles or CI/CD changes.

I wonder if superpowers should be conditionally used for bigger features, as it's rather a waste for small bugfixes or config toggles or CI/CD changes.

schaefera added 1 commit 2026-10-29 08:13:34 +00:00

CLAUDE.md: superpowers per session, for bigger features only (PR comment 2265) ... e6f87a89b3

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

claude-bot commented 2026-10-29 08:13:42 +00:00

Author

Collaborator

• Copy link
• Copy Markdown

Agreed. This PR already sets it up that way: superpowers stays off in .claude/settings.json, and a session for a bigger feature turns it on at launch:

claude --settings '{"enabledPlugins":{"superpowers@claude-plugins-official":true}}'

Settings passed at launch take precedence over the project's. I checked: with this flag, a session in this repo lists all 14 superpowers:* skills, and without it, none. e6f87a8 documents this in CLAUDE.md ("Keeping context small"). By "bigger feature" it means new F-feature work, a multi-PR issue or a design change, not fixes, config toggles or CI changes. A shell alias (e.g. claude-big) would save typing the flag. Tell me if you want one in your dotfiles.

This also settles the review's first low: small PRs simply go without the superpowers debugging and review skills.

Agreed. This PR already sets it up that way: superpowers stays off in `.claude/settings.json`, and a session for a bigger feature turns it on at launch: ``` claude --settings '{"enabledPlugins":{"superpowers@claude-plugins-official":true}}' ``` Settings passed at launch take precedence over the project's. I checked: with this flag, a session in this repo lists all 14 `superpowers:*` skills, and without it, none. e6f87a8 documents this in CLAUDE.md ("Keeping context small"). By "bigger feature" it means new F-feature work, a multi-PR issue or a design change, not fixes, config toggles or CI changes. A shell alias (e.g. `claude-big`) would save typing the flag. Tell me if you want one in your dotfiles. This also settles the review's first low: small PRs simply go without the superpowers debugging and review skills.

schaefera approved these changes 2026-10-29 08:17:47 +00:00

claude-bot changed title from WIP: Claude Code config: smaller start-up context for this repo to Claude Code config: smaller start-up context for this repo 2026-10-29 08:18:03 +00:00

claude-bot merged commit b3dba38b6a into main 2026-10-29 08:34:35 +00:00

claude-bot referenced this pull request from a commit 2026-10-29 08:34:35 +00:00

Merge pull request 'Claude Code config: smaller start-up context for this repo' (#110) from chore/claude-config into main

Sign in to join this conversation.

Reviewers

No reviewers

schaefera

Labels

Clear labels   

priority

Could

MoSCoW: nice to have, done when time allows

  

priority

Must

MoSCoW: the release isn't viable without it

  

priority

Should

MoSCoW: important, but the release works without it

  

priority

Won't

MoSCoW: not this release (kept for later)

  

type

Bug

No labels

priority

Could

priority

Must

priority

Should

priority

Won't

type

Bug

Milestone

Clear milestone

No items

No milestone

Projects

Clear projects

No items

No project

Assignees

Clear assignees

No assignees

3 participants

Notifications

Subscribe

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference

schaefera/utStudyBIble!110

No description provided.