This is my public, version-managed library of AI-agent skills.
Current release: V0.9.2. The skills now share one small default set, a four-theme catalogue, and on-demand loading across supported agents.
The skills are being adopted and adapted from my experience using AI agents to build software. They are not a neutral collection of generic prompts. Each skill reflects how I want an agent to think, plan, implement, review, and learn while working with me.
Using every available skill by default creates two problems:
- The agent spends too much context carrying instructions that do not apply to the current task.
- Skills change over time, so keeping many installed copies updated becomes manual work.
This repository is one canonical, version-controlled source for the skills I use across Claude Code, Codex, Pi, and Gemini via Antigravity. Author installs link the default skills from the checkout; consumer plugin installs expose the three default skills and package the full library as data. Ask-kb uses the generated themed catalogue to load a selected skill only when needed, so there is one copy to edit and one copy to update.
The second problem is handled by bin/skills-sync, which installs the library into every agent, keeps those installs reconciled with the repository, and gates releases. Git hooks run it, so a skill added in a commit reaches every agent without anyone remembering a step.
git clone https://github.com/kiran-brahma/personal-ai-skills.git
cd personal-ai-skills
bin/skills-sync bootstrap --role author --applybootstrap asks what kind of machine this is. An author machine installs from its own working tree and wires the git hooks, so an edit is live in the next session. A consumer machine installs the published plugin from the marketplace and tracks releases instead, so a local edit never moves its install.
Setting core.hooksPath is the only manual step in the whole setup. Everything after it is automatic.
Agents read skills at startup, so restart them after installing.
| Agent | Reads from | Invocation |
|---|---|---|
| Claude Code | ~/.claude/skills |
/skill-name |
| Codex | ~/.agents/skills |
$skill-name, or the /skills selector |
| Pi | ~/.pi/agent/skills |
/skill:name |
| Antigravity | ~/.gemini/config/skills |
/<skill-name> |
Codex and Pi both read ~/.agents/skills, the cross-harness Agent Skills location. Pi deduplicates by real path, so the overlap costs nothing.
The sections above are for my own machines. To use these skills yourself, fork the repository and install only what you need. Everything you don't install stays in your fork, and Ask-kb can still find and run it.
-
Fork
kiran-brahma/personal-ai-skillson GitHub, then:git clone https://github.com/<you>/personal-ai-skills.git ~/.kb-skills ~/.kb-skills/bin/skills-sync setup
setupadds this repository as theupstreamremote, installs the three default skills into every supported agent it finds, and wires the git hooks. -
Restart your agent and ask Ask-kb to set up optional skill installs. It loads
setup-skills, which recommends optional profiles (coding,typescript,deep-review,business, and others inprofiles.json). Your choice lives inmy-skills.txt, per machine and git-ignored. Change it any time through Ask-kb orbin/skills-sync select --add profile:business. -
Pull my releases with
bin/skills-sync sync. It merges the latest release tag, never unreleased work, and reinstalls. New skills arrive uninstalled but reachable.
Your own skills belong in your fork. When no skill fits a repeated need, Ask-kb loads decide-skills to help build one: bin/skills-sync new <name> --library, then select --add <name> to install it.
Use the default set. The plugin installs Ask-kb, Grilling, and Implement with no selection. It also packages the complete repository so Ask-kb can read on-demand skills without exposing their descriptions at startup:
claude plugin marketplace add kiran-brahma/personal-ai-skills && claude plugin install kb@kb-skills
codex plugin marketplace add kiran-brahma/personal-ai-skills && codex plugin add kb@kb-skillsInvoke plugin skills as /kb:skill-name in Claude Code. Choose either the plugin or the global installation on a machine so the same skills are not exposed twice.
You do not need to memorise anything:
- Any skill-driven request: Ask-kb lists every skill in the relevant theme, recommends one, and waits for your choice before loading it. Cross-theme requests get one combined list.
- Direct skill use: name the skill or select it from Ask-kb. Grilling remains available directly. Implement runs only after you name it or choose it from the menu.
- Find or build a skill: ask Ask-kb to search the catalogue or help create a skill in your fork.
These are personal skills, not a neutral toolkit. They encode how I want an agent to work with me. Several name me directly and are tuned to my writing and my blogs: cognitive-editor, content-fence, and musings-reviewer. Most of the engineering skills are portable, but if you want a library shaped around your own judgment, fork this and adapt it rather than installing it as-is. That is what I did with the sources below.
Much of the value is not mine. The engineering and productivity workflows are adapted from Matt Pocock's skills, the supporting lenses from pstack and thermos, and the business methods from gstack. Those authors distribute their own work, kept current by them, and you may prefer to take it from the source. Licences and attribution are in THIRD_PARTY.md; local adaptations and pinned upstream commits are recorded in registry.yaml.
bin/skills-sync doctor # what is out of sync, changing nothing
bin/skills-sync refresh # reconcile every agent with the repository
bin/skills-sync validate # check the library against the layout contractrefresh only manages symlinks that point into this repository. A real directory, or a link into another skill library, is reported as a conflict and never touched, so it cannot delete skills it did not install.
The git hooks make this automatic: pre-commit validates, and post-commit, post-merge, and post-checkout reconcile the install.
I start with well-written external skills, then adapt them to my own working style. A local change is intentional when it improves the way I work, even when the upstream skill is already strong.
The repository maintains one canonical local version of each adopted skill. It does not maintain separate Claude, Codex, Pi, or Antigravity copies. Harness-specific discovery and installation details belong in adapters/, while the workflow itself remains portable.
Upstream updates are inputs for review, never applied automatically:
bin/skills-sync upstream # what the original authors changed since the pinned commit
bin/skills-sync adopt --apply # three-way merge, preserving local adaptationsupstream reports what moved and, more usefully, where an upstream change touches a file I have also edited. adopt merges the rest and leaves conflict markers where both changed the same lines, because that is a judgment no tool should make. An unresolved merge fails validate, so it cannot reach a commit or a release.
Before adopting an update I compare:
- The new upstream behavior.
- The local adaptation and the reason it exists.
- The evaluation cases and observed behavior.
- The effect on other skills and the overall workflow.
The registry records the upstream repository, commit, license, local adaptations, status, and update policy for each source. upstream --pin moves that commit only after the review is done.
One engineering workflow, based on Matt Pocock's skills, is:
grill-with-docs → to-spec → goldilocks-review → to-tickets → implement → code-review
The workflow is deliberately gated. After a PRD or spec is complete, goldilocks-review checks for a solution that is simple in the resulting system, not merely easy for an AI to generate. It gives minimal coupling and operational simplicity priority, with a preference for deep modules behind limited interfaces.
For a major code change, the sequence is:
implement → thermos → fix findings → code-review → pull request
Every pull request must complete code-review. Detailed security audits use the separately loaded Codex Security workflow rather than the normal coding workflow.
Only three skills are installed by default: Ask-kb, Grilling, and Implement. Ask-kb uses four themes to organize the full skill catalogue. Each skill has exactly one primary theme; themes do not install their skills.
The default set stays small as the library grows. Each installed skill's description sits in the agent's context from startup, and Codex silently truncates once the combined descriptions pass a fixed budget. The library is split:
- Installed (
skills/): Ask-kb routes requests, Grilling handles design interviews, and Implement is the explicitly chosen implementation workflow. - On demand (
library/): all other skills. Agents do not load their descriptions at startup. Ask-kb reads the generated themed catalogue and loads a selected skill's full instructions.
The four themes are Coding, Writing, Productivity, and Misc. Optional install profiles remain separate bundles for users who want more skills installed on a particular machine. The index groups by theme, not profile. Both tiers use the same flat, one-level layout. Claude Code searches only one level, and Pi stops at the first SKILL.md it finds. bin/skills-sync validate enforces the layout, checks that each skill has one theme and the catalogue is current, and charges only the installed tier against the budget. policies/skill-format.md records which agent imposes which constraint.
Move a skill between tiers with bin/skills-sync tier <name> library or bin/skills-sync tier <name> installed. Ask Ask-kb to load skill-tiers when changing repository defaults.
- Matt Pocock: foundation engineering and productivity workflows.
- pstack: targeted supporting lenses for architecture, codebase understanding, blast-radius analysis, adversarial review, verification, technical writing, and long-running work.
- Thermos: deep review for major changes before Matt's final code review.
- Continual learning: proposal-first updates to
AGENTS.mdandCLAUDE.mdbased on durable lessons. - gstack: selected product-discovery and founder-review methods, adapted into portable business workflows.
- HyperFrames: not vendored.
decide-skillsruns it from heygen-com/hyperframes when Ask-kb routes a video request there. - Codex Security: external-only workflow for detailed security audits.
The source list and pinned references are in registry.yaml. Adapted-license notices are in THIRD_PARTY.md. Upstream provenance documents that are not skills live in packages/.
bin/skills-sync new my-skill -d "What it does. When it fires. Not for X, use Y." # installed
bin/skills-sync new my-skill -d "..." --library # on demand; add it to a profile, then `index`The scaffold refuses an over-budget description and a name that will not resolve. A description is a trigger contract, not a summary: it answers whether the skill should fire, and the detail belongs in the body where it costs nothing until the skill is invoked. The rule is in policies/skill-format.md.
After writing the body, record the skill in registry.yaml and run bin/skills-sync validate.
bin/skills-sync release # preflight only
bin/skills-sync release --apply # publish, then verify the install movedPreflight blocks on a dirty tree, a detached HEAD, a validation failure, an invalid plugin manifest, a tag that already exists, and a missing changelog entry. Only then does it write the version, commit, tag, push, refresh the local install, and re-check that the install actually moved.
VERSION holds the major and minor version. The patch is derived from the commits touching skills/ or library/ since the last tag, so it cannot fail to move when the library does. Claude Code and Codex ship an update only when the manifest version changes, which makes a forgotten bump a silent failure rather than a loud one.
personal-ai-skills/
├── AGENTS.md Canonical repository control document
├── CLAUDE.md Claude entrypoint that refers to AGENTS.md
├── GEMINI.md Gemini entrypoint that refers to AGENTS.md
├── SKILLS.md Top-level catalogue: which tier, which router
├── VERSION Major and minor version; the patch is derived
├── bin/skills-sync Validation, install, release, and upstream tooling
├── .githooks/ Hooks that validate and reconcile on every commit
├── .claude-plugin/ Claude Code plugin and marketplace manifests
├── .codex-plugin/ Codex plugin manifest
├── skills/ Default-loaded Ask-kb, Grilling, and Implement
├── library/ All other skills, loaded on demand through Ask-kb
├── adapters/ Agent-specific discovery and installation details
├── packages/ Provenance documents for vendored upstream packages
├── policies/ Repository rules and decision records
├── evals/ Behavioral evaluation cases
├── registry.yaml Skill inventory and provenance metadata
├── profiles.json Four theme maps plus optional install bundles
├── THIRD_PARTY.md Notices for adapted external skills
├── CHANGELOG.md Release index
└── changelog/ Release notes, one file per month
The GitHub repository is the canonical source. Local copies are for development and evaluation; agents should consume the committed and pushed version.