Set Up Private Context Files for Coding Agents
Safely evaluates and adopts Context Kit's personal-context files, inspecting installers before running and protecting private data.
Why it matters
The user hires this asset to safely organize and maintain personal context files (identity, preferences, mental models, protocols) that coding agents can read at session start, without leaking sensitive data into public repos or third-party tools.
Outcomes
What it gets done
Inspect and safely run the Context Kit installer after reviewing source code and target paths
Create local-first Markdown artifacts for voice, mental models, and protocols in gitignored directories
Prevent secrets, API keys, and private relationship data from being stored in agent context files
Establish a review cadence to keep personal context current and remove stale assumptions
Install
Add it to your toolbox
Run in your project directory:
curl -fsSL https://spark.entire.vc/get/ag-context-kit | bash Overview
Context Kit
This skill helps a user safely adopt Context Kit's personal-context Markdown artifacts - inspecting the installer before running it, choosing a storage location, creating a minimal starter set, and applying safety rules around secrets and private data. Use it when setting up personal context files for a coding agent or reviewing whether a one-line installer is safe to run - it's for local personal context hygiene, not a secrets manager or CRM replacement.
What it does
Helps a user safely adopt Context Kit, an external project that organizes personal context into Markdown artifacts and companion coding-agent skills, deciding what to adopt, where to store it, and how to avoid turning a useful context system into a pile of sensitive data.
git clone https://github.com/JDDavenport/context-kit.git
cd context-kit
sed -n '1,220p' scripts/install.sh
find templates skills -maxdepth 2 -type f | sort
Every personal context file is treated as private by default - these files can hold identity details, family context, work history, contact notes, mental models, health constraints, or relationship information, never pasted into third-party tools, public repos, issue trackers, or model contexts without explicit user approval of the exact subset. Safety rules require inspecting a remote install script (command, source repo, target paths) before running it - cloning the repository and reading scripts/install.sh directly rather than piping curl straight to a shell; never storing passwords, API keys, recovery codes, private keys, session tokens, or payment details in context artifacts; storing only plaintext-comfortable information in CRM/contact notes; confirming .gitignore excludes the chosen private-context directory before adding files to a repo; and separating personal from company-confidential or customer-confidential information when adapting Context Kit to a team setting.
The setup workflow asks what the user wants improved (session startup context, voice consistency, relationship memory, open-loop tracking, daily briefings, handoff summaries), inspects the upstream project and installer first, chooses a storage location (a default location under the agent's own context directory, a project-local ignored directory, or a portable private notes repo with explicit sync rules), creates a minimal starter set (pca-wiki.md for durable identity/domains, pca-mental-models.md for decision rules, pca-voice.md for writing preferences, pca-protocols.md for hard rules/boundaries) rather than filling everything at once, adds only enough detail to make the next session useful, and sets a recurring review cadence since stale personal context is worse than no context when it drives decisions.
Before running any installer, an installation review checklist verifies the repository URL matches what the user intended, the script only writes to expected local directories, it doesn't upload files/send telemetry/edit shell startup files unexpectedly, target directories aren't inside a public repo, and the user has a rollback path (removing the copied templates and skills).
When to use - and when NOT to
Use it when a user wants to set up durable personal context files, compare Context Kit's pattern against an existing memory/notes system, adapt a template structure without pasting private details into chat, review whether a one-line installer is safe to run, or plan safer local-first storage for CRM notes, open loops, or briefings. It doesn't verify upstream license or installer behavior on its own - inspect the live repository first - and it isn't a substitute for a dedicated secrets manager, CRM, password vault, or medical record system, nor is it for collecting private information about other people without a legitimate reason.
Inputs and outputs
Input is the user's context-management goal and, when installing, the Context Kit repository to inspect. Output is a reviewed, minimal starter set of local context files (or a clear list of what an installer would write) plus a .gitignore entry protecting the private context directory.
Integrations
Wraps the external JDDavenport/context-kit project - its templates, skills, and installer script - with an inspect-before-install safety layer.
Ongoing best practices for maintaining context files: keep them short enough that an agent can read them at session start without drowning in stale detail; separate durable facts from temporary state, using a project workplan or task tracker for the latter; label assumptions and uncertain memories instead of presenting them as facts; review personal context after major life, role, health, or project changes; and store voice examples and anti-examples separately from private identity details when possible.
Who it's for
Users who want durable personal context for their coding agent sessions - identity, mental models, voice, protocols - without blindly running a remote installer or accidentally committing sensitive personal data to a repository.
Source README
Context Kit
When to Use
Use this skill when the user wants to:
- Set up durable personal context files for Claude Code or another coding agent
- Compare Context Kit's Personal Context Artifact pattern with an existing memory or project-notes system
- Adapt a context template structure without copying private details into the chat
- Review whether a one-line installer or downloaded skill pack is appropriate before running it
- Create a safer, local-first plan for CRM notes, open loops, session digests, or morning briefings
Overview
Context Kit is an external project that organizes personal context into Markdown artifacts and companion
Claude Code skills. This skill helps the user decide what to adopt, where to store it, and how to avoid
turning a useful context system into a pile of sensitive data.
Treat every personal context file as private by default. These files may contain identity details, family
context, work history, contact notes, mental models, health constraints, or relationship information. Do
not paste them into third-party tools, public repositories, issue trackers, or model contexts unless the
user explicitly approves the exact subset.
Safety Rules
- Do not run a remote install script until the user has seen the command, source repository, and target
paths it will write to. - Prefer cloning or downloading the repository for inspection before executing any installer:
git clone https://github.com/JDDavenport/context-kit.git cd context-kit sed -n '1,220p' scripts/install.sh - Never store passwords, API keys, recovery codes, private keys, session tokens, or payment details in
personal context artifacts. - If the user wants contact notes or CRM files, store only information they are comfortable keeping in
local plaintext Markdown. - Before adding these files to a repo, confirm
.gitignoreexcludes the chosen private context directory. - If adapting Context Kit to a team or company setting, separate personal context from company-confidential
or customer-confidential information.
Setup Workflow
- Ask what the user wants Context Kit to improve: session startup context, voice consistency, relationship
memory, open-loop tracking, daily briefings, or handoff summaries. - Inspect the upstream project and installer before running anything.
- Choose a storage location:
- Claude Code default:
~/.claude/context/and~/.claude/skills/ - Project-local context:
.agent/context/or another ignored directory - Portable setup: a private notes repo with explicit sync rules
- Claude Code default:
- Create a minimal starter set before filling everything:
pca-wiki.mdfor durable identity and domainspca-mental-models.mdfor decision rulespca-voice.mdfor writing preferencespca-protocols.mdfor hard rules and boundaries
- Add only enough detail to make the next agent session useful. Leave sensitive, speculative, or outdated
details out until there is a clear reason to include them. - Add a recurring review cadence. Personal context goes stale quickly; stale context is worse than no
context when it drives decisions.
Installation Review Checklist
Before running an installer, verify:
- The repository URL is exactly the one the user intended
- The script writes only to expected local directories
- The script does not upload files, send telemetry, or edit shell startup files unexpectedly
- The target directories are not inside a public repo
- The user has a rollback path, such as removing the copied templates and skills
If anything is unclear, stop at inspection and provide the user with the exact lines that need review.
Examples
Example: Inspect before installing
git clone https://github.com/JDDavenport/context-kit.git
cd context-kit
sed -n '1,220p' scripts/install.sh
find templates skills -maxdepth 2 -type f | sort
After inspection, summarize the files that would be installed and ask for confirmation before running the
installer.
Example: Create a private project-local context directory
mkdir -p .agent/context
printf '.agent/context/\n' >> .gitignore
cp ~/Downloads/context-kit/templates/pca-wiki.md .agent/context/pca-wiki.md
Then trim the template to the minimum useful fields for the project instead of filling every personal
section immediately.
Best Practices
- Keep context files short enough that an agent can read them at session start without drowning in stale
detail. - Separate durable facts from temporary state. Use project workplans or task trackers for temporary state.
- Label assumptions and uncertain memories instead of presenting them as facts.
- Review personal context after major life, role, health, or project changes.
- Store voice examples and anti-examples separately from private identity details when possible.
Common Pitfalls
- Running a shell installer directly from
curlwithout inspecting it first - Committing personal context files to a public repository
- Storing secrets because "the agent needs to know everything"
- Letting relationship or health notes become outdated and still treating them as current
- Copying upstream paid or license-unclear content instead of linking to it or writing original local notes
Limitations
- This skill does not verify the current upstream license or installer behavior on its own; inspect the live
repository before running commands. - It does not replace a dedicated secrets manager, CRM, password vault, or medical record system.
- It is for local personal context hygiene, not for collecting private information about other people
without a legitimate reason.
FAQ
Common questions
Discussion
Questions & comments · 0
Sign In Sign in to leave a comment.