Personal Agent Configuration

What this is

A personal installer that distributes my agent rules, commands, and skills to the local configuration of every supported agent. It is a personal policy repository, not a curated public list: the content encodes my preferences and is published so that it can be installed and updated with one command.

• Audience: primarily me; usable by anyone who wants opinionated, versioned agent rules.
• Scope: local distribution only. The installer never fetches rules from elsewhere; it copies the content committed in this repository.
• Status: personal policy. External material is adapted selectively and recorded in PROVENANCE.md.

Install / update the AI rules

Distributes the rules in docs/, plus commands/ and skills/, to your local agent configurations:

Agent	Rules	Commands	Skills
OpenCode V2	~/.config/opencode/AGENTS.md	~/.config/opencode/commands/	reads ~/.agents/skills/
Claude Code	~/.claude/CLAUDE.md	~/.claude/commands/	~/.claude/skills/
Codex	~/.codex/AGENTS.md	~/.codex/prompts/	~/.agents/skills/
Pi	~/.pi/agent/AGENTS.md	~/.pi/agent/prompts/	reads ~/.agents/skills/

Rules are concatenated into each agent's native single-file location. Skills are written to only two directories — ~/.agents/skills/ (read by OpenCode V2, Codex, and Pi) and ~/.claude/skills/ (read by Claude Code) — because each harness scans several skills directories at once and duplicate skill names across them collide. OpenCode V2 officially supports ~/.agents/skills/ as a global compatibility source; using it avoids a second copy in its native skills directory.

Requires git and Python 3.

Release install (recommended)

Every release publishes a source archive and its SHA-256 checksum under Releases. Set VERSION to the release tag you want (for example v1.0.0), then download, verify the checksum, and extract.

macOS / Linux:

VERSION=v1.0.0
BASE="https://github.com/doberkofler/personal-agent-config/releases/download/${VERSION}"
TMP="$(mktemp -d)"
curl -fsSL -o "$TMP/personal-agent-config-${VERSION}.tar.gz" "$BASE/personal-agent-config-${VERSION}.tar.gz"
curl -fsSL -o "$TMP/SHA256SUMS" "$BASE/SHA256SUMS"
( cd "$TMP" && sha256sum -c SHA256SUMS )      # macOS: shasum -a 256 -c SHA256SUMS
mkdir -p ~/.ai-rules/src
tar -xzf "$TMP/personal-agent-config-${VERSION}.tar.gz" -C ~/.ai-rules/src --strip-components=1
python3 ~/.ai-rules/src/sync_ai_rules.py

Windows (PowerShell):

$Version = "v1.0.0"
$Base = "https://github.com/doberkofler/personal-agent-config/releases/download/$Version"
$Tmp = Join-Path $env:TEMP "ai-rules-$Version"
New-Item -ItemType Directory -Force -Path $Tmp | Out-Null
Invoke-WebRequest "$Base/personal-agent-config-$Version.tar.gz" -OutFile "$Tmp/personal-agent-config-$Version.tar.gz"
Invoke-WebRequest "$Base/SHA256SUMS" -OutFile "$Tmp/SHA256SUMS"
Get-Content "$Tmp/SHA256SUMS"   # confirm it matches Get-FileHash -Algorithm SHA256
New-Item -ItemType Directory -Force -Path "$HOME/.ai-rules/src" | Out-Null
tar -xzf "$Tmp/personal-agent-config-$Version.tar.gz" -C "$HOME/.ai-rules/src" --strip-components=1
python "$HOME/.ai-rules/src/sync_ai_rules.py"

Or clone at a pinned tag

VERSION=v1.0.0
git clone --branch "$VERSION" --depth 1 https://github.com/doberkofler/personal-agent-config.git ~/.ai-rules/src
python3 ~/.ai-rules/src/sync_ai_rules.py

Local development

If you are developing the rules rather than consuming them, work in a main checkout and run the installer directly against your working tree:

git clone https://github.com/doberkofler/personal-agent-config.git
cd personal-agent-config
python3 sync_ai_rules.py --dry-run   # preview the plan
python3 sync_ai_rules.py             # install the working-tree content

When the script runs from a checkout it uses that directory as its source. It skips git pull when the checkout has local changes (and when it is checked out at a fixed revision), so you can edit docs/, commands/, and skills/ and re-run to install your edits.

Update and verify

python3 ~/.ai-rules/src/sync_ai_rules.py            # install the checked-out content
python3 ~/.ai-rules/src/sync_ai_rules.py --check    # revision status and content integrity

--check reports two independent statuses: the installed revision against upstream, and every installed file against its recorded SHA-256 (modified, missing, unexpected, obsolete). It exits nonzero on content drift or when the manifest cannot be read; the repository status is informational, since a release archive has no git metadata. To update, install a newer release tag, or git -C ~/.ai-rules/src fetch --tags && git -C ~/.ai-rules/src checkout "$VERSION".

Optional alias: alias ai-rules='python3 ~/.ai-rules/src/sync_ai_rules.py'.

Trust and update model

Installing runs code from this repository. Releases are built by CI from an annotated v* tag and published with a SHA-256 checksum. Verifying the checksum confirms the downloaded archive matches the published artifact; it does not protect against a compromised GitHub account, a malicious maintainer, or a tampered release. Pin an explicit VERSION, download only over HTTPS from this repository, and review the script before running it.

Notes

Maintainers cut a release by pushing a v* tag; CI runs release.py, which builds dist/personal-agent-config-<tag>.tar.gz and dist/SHA256SUMS and attaches both to the GitHub release.

See PROVENANCE.md for content origins, intentional adaptations, upstream maintenance, distribution behavior, and the release process.

Verification

This repository distributes documentation. It has no Node toolchain and does not compile, lint, or test the TypeScript examples. Verify a change with:

python3 -m unittest discover -s tests -t .
git diff --check
python3 sync_ai_rules.py --check

Distribution integrity must report ok.

What gets installed

Rules (docs/)

Concatenated in filename order into each agent's global rules file.

Document	Focus
general-guidelines.md	Persona, clarification protocol, verbosity, validation sources
lint-rules.md	Oxlint static-analysis and Oxfmt formatting configuration
mui-rules.md	Material UI usage, theming, and color schemes
react-rules.md	React component and hook design
sql-plsql-rules.md	SQL and PL/SQL coding standards
typescript-rules.md	Canonical TypeScript baseline and Zod validation

Commands (commands/)

OpenCode-style command protocols, distributed to every supported agent's command directory.

Command	Purpose
dev-decompose.md	Decomposes a researched plan into wave-ordered, self-contained task specs
dev-find-reinventions.md	Finds internal duplication and reinvented builtin/framework/npm functionality
dev-reset.md	A formalized epistemic reset protocol
dev-verimode.md	A verification-first, documentation-grounded response protocol

Skills (skills/)

Each skill is a directory containing a SKILL.md; the installer copies the whole directory.

Skill	Purpose
code-review	Independent review of a branch, commit range, or worktree against repository standards
codebase-design	Vocabulary for designing deep modules and choosing seams
diagnosing-bugs	Diagnosis loop for hard bugs and performance regressions
domain-modeling	Build and sharpen a project's domain model and ADRs
grill-me	Relentless interview to sharpen a plan or design
grill-with-docs	Relentless interview that also writes ADRs and a glossary
grilling	Grill a plan, decision, or idea to stress-test it
handoff	Compact a conversation into a handoff document
oracle-local	Query, test, and validate against a local Oracle database
prototype	Build a throwaway prototype to answer a design question
research	Investigate a topic against primary sources and capture the findings
tdd	Test-driven development and red-green-refactor
wait-what	Re-pitch the previous message when it did not land
writing-for-agents	Write skills, AGENTS.md, and CLAUDE.md for agents

General meta-prompts

Short prompts used while working with an agent. Not distributed; kept here for reference.

Name	Description	Prompt
Epistemic reset	Forces separation of fact from inference	“What do you actually know vs. assume here?”
Assumption audit	Surfaces hidden priors before they compound	“List every assumption embedded in your last response.”
Steelman the opposite	Breaks confirmation bias	“What’s the strongest case against your current approach?”
Constraint surfacing	Catches over-engineering and scope creep	“What are you optimizing for, and did anyone ask you to?”
Uncertainty declaration	Prevents hallucination laundering	“Rate your confidence per claim and flag what you’d need to verify.”

External resources

Inclusion criteria: the link must be a general reference or collection for a topic covered by docs/, commands/, or skills/, and it must be non-commercial in intent. One-off blog posts, SEO prompt dumps, and personal marketing pages are excluded.

OpenCode

• awesome-opencode - a curated list of OpenCode resources.

Prompt engineering

• prompts.chat - share, discover, and collect prompts.
• awesome-prompts - a curated list of prompts.

Cursor rules

• awesome-cursorrules - configuration files with custom rules and behaviors.
• cursor-rules - Cursor directory rules.

Frontend design

• impeccable - curated anti-patterns for impeccable frontend design.