Manage · Memory setup prompt
Seed repository memory without erasing its history
The Memory setup prompt tells a coding agent how to inspect a repository, seed an inherited engineering baseline, write repository-specific memory beneath it, and report what changed. You still review the agent's diff before keeping it.
For: the repository owner using a coding agent to create or repair the brain layer while preserving existing project knowledge
What it does, and why it helps
The served feature is one Markdown file, .claude/brain/brain-bootstrap-prompt.md. Open Memory setup prompt, copy the prompt block, and give it to a fresh coding-agent session with read and write access to the target repository. The prompt asks the agent to inspect the real stack, commands, routes, features, and data model before writing. It tells the agent to merge or skip existing files, never overwrite them, and to name each choice in its final report.
The prompt separates an inherited baseline from facts learned in the target repository. If starter-knowledge-pack/ is available, it names five pack inputs for principles, patterns, anti-patterns, errors, and lessons. Repository-specific learning belongs below a divider. If the pack is absent, the prompt asks the agent to say so and write a smaller baseline from named principles. This is an instruction to an external agent, not a verified installer. For initial RepoOps integration, use the reviewed repository setup flow instead.
The pain. A new repository starts with no durable explanation of its structure, decisions, or working rules. An empty template moves the problem forward without giving the next session useful context.
The point of view. Start with a clearly fenced inherited baseline, then add only facts the agent can find in the repository. Never let generic guidance pass as repository evidence, and never let a bootstrap overwrite reviewed state.
What gets easier. Giving a fresh agent one bounded brief for orientation, inherited rules, repository-specific files, session habits, and a final file-by-file report.
When it helps. Use the prompt when you deliberately want an AI-assisted brain layer in a new repository, or when an existing brain needs a careful merge rather than an automated replacement.
Its limits. The page only renders text. It does not run the prompt, verify the agent's inspection, test the generated files, install capture, enable team delivery, add the full scripts and CI tree, or predict the selected model's cost. The deterministic setup flow is separate.
Understand it in 30 seconds
Read the narration
- 0:00 A new repository has no memory of the decisions its next session needs.
- 0:06 Start with inherited discipline, then separate it from facts found in this repository.
- 0:13 The prompt asks the agent to inspect the stack and preserve existing files.
- 0:17 Its report separates inherited content from repo facts.
- 0:23 Review the diff before keeping it.
- 0:25 A prompt is not a verified install.
Synthetic example. Read the guide
Where to find it
Where to find it
- Desktop:
localhost:4000, then Local settings in the sidebar, then Memory setup prompt under Workspace utilities. - Keyboard: ⌘ K, then type “Memory setup prompt”.
When to use it
Start with the inherited pack available
Situation. The target repository has no brain layer, and you can copy starter-knowledge-pack into its .claude/brain directory or give the agent read access to a donor repository that contains it.
What you do. Open Memory setup prompt, copy the prompt block, paste it into a fresh agent session, and let the agent inspect before writing. Review the proposed tree and the distinction between inherited and repository-specific content.
What you see. The agent's final report should name every file it created, merged, or skipped. Pack-derived prose should appear above a divider, with target-repository facts below it.
What it establishes. You can review whether the baseline stayed generic and whether each repository claim is grounded. The prompt alone does not establish that the resulting memory is correct.
Add memory without clobbering existing guidance
Situation. The repository already has CLAUDE.md or some brain files, and their reviewed content must survive.
What you do. Keep the prompt's Step 0 instruction intact: do not overwrite existing files. Ask the agent to merge or skip each conflict and say which choice it made.
What you see. The final report lists preserved files and merged sections. For a repository with more than 100 pull requests, log directories begin forward from the current day rather than being reconstructed from old history.
What it establishes. The diff makes preservation decisions visible for human review. An unchanged old file is not proof that its contents are current; review it separately.
Before you start
- Supported versions
- RepoOps desktop v0.3.1, the release this guide was read against.
- Where it runs
- Local: Local settings, Workspace utilities, Memory setup prompt, when the selected repository contains .claude/brain/brain-bootstrap-prompt.md. There is no committed public HTML twin; RepoOps renders that Markdown source at request time. Hosted: Help and the owner/admin Prompts page link to a bundled copy in the shared document viewer.
- Permissions
- Reading locally uses the desktop operator surface. The hosted viewer needs an authenticated, entitled team member; the Prompts index is owner/admin only, while Help and a direct bundled document link are available to entitled members. Running the prompt needs whatever filesystem and version-control rights the chosen coding agent receives in the target repository.
- Connections
- A fresh coding-agent session must be able to read the target repository and write the reviewed scope. For the full inherited baseline, copy starter-knowledge-pack into the target brain or give the agent read access to a donor repository that contains it.
- Plan
- The local prompt has no plan gate. Hosted reading follows the hosted dashboard subscription and seat checks. The external coding agent and its model account are outside this feature's plan controls.
Configure it
- Choose the reviewed path first.
For initial local-app integration, prefer the deterministic repository setup flow. Use this prompt only when you want an external agent to inspect and author the brain content.
- Make the inherited source explicit.
Copy starter-knowledge-pack into the target brain, or give the agent read access to a donor copy. If neither is available, keep the prompt's fallback instruction and require the agent to disclose that the pack was absent.
- Paste into a fresh session.
A clean context keeps the target repository and this prompt at the center of the task. Do not remove Step 0: Adapt to this repo, because that is the evidence boundary between generic guidance and repository facts.
- Review the tree and diff.
Check every created, merged, and skipped file. Confirm that inherited sections stay fenced, repository facts match source, existing instructions survived, and empty-repo placeholders are marked TODO.
| Setting | Where | A sensible choice | Why it matters |
|---|---|---|---|
starter-knowledge-pack/ | Before you paste and Step 1 | Copy it into .claude/brain, or provide a readable donor copy | It supplies the named inherited baseline. Without it, the prompt asks the agent to disclose the absence and write the smaller fallback. |
Existing files | Step 0, Adapt to this repo | Merge or skip, never overwrite | Reviewed repository guidance must survive bootstrap; the agent must report its choice. |
Empty repository | Step 0, Adapt to this repo | Use clearly marked TODO placeholders | The prompt forbids presenting missing repository facts as known state. |
High-history repo | Step 0, Adapt to this repo | At more than 100 pull requests, start log directories forward from today | The prompt asks for current state without reconstructing old pull-request and session logs. |
Scope | Opening and Step 5 | Brain layer only | Skills, scripts, and CI gates belong to Repo setup prompt, not this prompt. |
What you should see
Inherited and repository-specific layers are distinct
Configuration. The starter pack is readable, existing files are preserved, and the agent completes Steps 0 through 5.
Expect. The prompt asks for inherited principles, patterns, anti-patterns, errors, and lessons above a divider, with the target repository's own learning below it.
Verify. Read the diff. Check that pack prose did not acquire project details and that each project claim can be traced to the target repository.
The no-pack fallback stays honest
Configuration. No starter-knowledge-pack or readable donor copy is available.
Expect. The prompt asks the agent to state that fact and build a smaller inherited baseline from the named principles and work habits, rather than leave the files empty.
Verify. The final report names the missing pack. Treat the generated fallback as agent-authored prose and review it line by line.
Existing state is preserved visibly
Configuration. CLAUDE.md or one or more requested brain files already exist.
Expect. The prompt directs the agent to merge or skip instead of overwrite, then list that decision in its report.
Verify. Compare the before and after files in version control. If the agent cannot show the preserved text and explain the merge, do not keep the change.
Data and cost
- What is captured
- Opening the page captures no repository data. Once pasted, the prompt asks the selected coding agent to inspect manifests, documentation, routes, features, the data model, and existing brain files, then write Markdown, JSONL, settings, and instruction files in the target repository.
- Who can see it
- The local prompt source is a repository file. The hosted copy is product content bundled with the dashboard, not a tenant brain snapshot. Files created in the target repository are visible to its filesystem users and, if committed, to that repository's version-control audience.
- How long it is kept
- Not provided. The prompt sets no retention window. Its output remains in the target repository until a person or repository process changes or removes it; version-control history follows that repository's own policy.
- What leaves the machine
- Reading the local page makes no model call. Pasting the prompt into a hosted or remote coding agent can send the prompt and repository context to that agent's provider, under that tool's configuration. RepoOps does not proxy that manual model session.
- What it costs
- RepoOps makes no paid model call to render or bundle the prompt. Any tokens, tool calls, or provider charges come from the external coding-agent session and are not capped or priced by this feature.
When the result differs
| Symptom | Likely cause | Next action |
|---|---|---|
| The agent cannot find starter-knowledge-pack. | The directory was not copied into the target brain and no readable donor repository was supplied. | Supply one of those sources, or continue only with the prompt's disclosed no-pack fallback and review the generated baseline closely. |
| Memory setup prompt has no local page for the selected repository. | The local renderer needs .claude/brain/brain-bootstrap-prompt.md in that repository. The current brain-only scaffolder does not install the prompt file. | Read the bundled hosted copy, or copy the reviewed prompt source into the target repository before using the local leaf. |
| The result contains generic claims about the repository. | The agent skipped Step 0 or lacked enough source material to ground repository-specific files. | Reject the diff. Re-run with repository read access and require file evidence for each state claim; in an empty repo, keep explicit TODO markers. |
| An existing brain file was replaced. | The agent did not follow the merge-or-skip instruction. | Do not keep the replacement. Restore the reviewed file through version control, then merge only the sections you accept. |
| The result has no skills, scripts, or CI gates. | Memory setup prompt is intentionally limited to the brain layer. | Use Repo setup prompt for the broader repository structure, or use the deterministic repository setup flow for initial integration. |
| The hosted viewer shows a product prompt even though no brain snapshot was published. | The four bootstrap documents use an allowlisted bundled fallback and are not read from a tenant snapshot. | This is expected. The viewer labels it A RepoOps bootstrap prompt bundled with the dashboard and names brain-bootstrap-prompt.md. |
- Disable
- Not provided. The prompt is inert text until someone gives it to an external agent.
- Roll back
- Not provided by the prompt. Review and revert its file changes through the target repository's version-control process. The separate deterministic integration has its own reviewed rollback path.
- Revoke access
- Not provided. The prompt creates no RepoOps credential or access grant. Revoke access in the external coding agent or repository system that received it.
- Delete
- Not provided. Remove accepted output files only through a reviewed repository change; the prompt does not track or delete what an external agent wrote.
Related tasks
Maintenance evidence
- Feature id
brain-bootstrap(spine leafbrain-bootstrap-prompt)- Owner
- Vince onboarding (P5, P9.B2), Guide: LDG-0717
- Supported product version
- RepoOps v0.3.1
- Last verified
- 2026-09-15, read against origin/main at 52366bb6d; labels read from the served tab source; prompt source, local renderer, hosted bundle, deterministic alternative, and tests read in the same pass
- Example fixtures
- lib/brain/bootstrap-detector.test.mjs, lib/install/scaffolder.test.mjs, lib/install/scaffold-into-repo.test.mjs, lib/install/repository-integration.test.mjs, lib/md-renderer.frontmatter.test.mjs, website/lib/render-brain-markdown.test.ts, scripts/gen-bootstrap-bundle.test.mjs
- Source references
lib/canonical-spine.json,lib/canonical-tabs.mjs,.claude/brain/brain-bootstrap-prompt.md,lib/routes/mirror-files.mjs,lib/md-renderer.mjs,website/app/team/(home)/help/page.tsx,website/app/team/(home)/prompts/page.tsx,website/app/team/(home)/docs/[file]/page.tsx,website/lib/bootstrap-docs-content.ts,scripts/gen-bootstrap-bundle.mjs,lib/install/scaffold-into-repo.mjs,lib/install/repository-integration.mjs,lib/kernel-file-set.mjs- Documentation review
- Independent review requested on the slice pull request; not yet recorded.
- Video review
- Narrated story rendered and published 2026-09-26 (render 5577838096b3, LDG-1018) with the breadcrumb Local settings, checked against main at 0f24215b5. One frame, the captions and the transcript were reviewed by the authoring agent, not an independent reviewer; the audio was not listened to by a person. Narration is the provisional Windows voice until LDG-0721.
Last updated