Quick Links: Website | Browse Skills | View Agents | Explore Personas | Check Maintenance Leads | About the author / hiring
Agent Skills is my shared toolbox and playground for agent work: reusable capabilities, bounded workers, persona contracts, and lifecycle hooks. I've spent hours re-solving problems I already solved once. This repo is here to stop that cycle by keeping useful agent patterns versioned, documented, and easy to find before I build something new from scratch.
Public repo, private runtime. The code, prompts, contracts, and docs are public, but some paths expect private infrastructure: memory services, browser bindings, model gateways, media storage, credentials, or agent homes. Treat this as a working blueprint and research playground, not a finished turnkey SDK. Many patterns copy cleanly; some commands need environment setup before they run in a fresh clone.
This repo is safe to show because it focuses on patterns: contracts, lifecycle hooks, maintainer reports, and reusable skill boundaries. Private runtime data, credentials, regulated context, and operational details stay outside the repo.
Pick the path that matches what you want to explore:
| If you want to... | Start here |
|---|---|
| Solve a task | skills/ for browser, memory, analysis, model, review, and media capabilities |
| Study agent behavior | agents/ for worker roles, boundaries, receipts, and stop conditions |
| Explore identities | personas/ for registry records, memory probes, and voice-readiness evidence |
| Borrow safety rails | hooks/ for quality gates, memory-first rules, and completion discipline |
| Find maintenance leads | reports/agent-maintainer/latest.md for triage notes and possible follow-up work |
| Know who built this | RESUME.md — the author's resume; this repo is the working evidence behind it |
# Find skill contracts
find skills -maxdepth 2 -name SKILL.md | sort
# Find agent contracts
find agents -maxdepth 2 \( -name AGENTS.md -o -name persona.yaml \) | sort
# Read the persona registry guide
sed -n '1,160p' personas/README.md
# Inspect the latest maintainer triage report
sed -n '1,180p' reports/agent-maintainer/latest.mdA few helpful habits:
- Use
README.mdfiles as friendly guides; useSKILL.mdfiles as operational contracts. - When a skill looks relevant, start with its
SKILL.md. That file explains what the capability promises and how it should be used. - Before writing a new utility, take a quick look through the skill list. There may already be a close starting point.
Deploy to configured agent homes:
# Deploy skills and hooks to local agent homes
./deploy.sh
# Preview what would change
./deploy.sh --checkI maintain (and sadly break) these kinds of projects and skills continuously. My apologies in advance for a static table. Here's my current focus:
Persona Dream Agentic memories to movies |
Surf ChatGPT browser control for agents |
Battle Red-Blue Team genetic fuzzing arena |
T'au Memory-First Zero-Trust Agent Harness |
Extractor Docs into hierarchical evidence trees |
Dogpile Search ArXiv, YT, GH, RSS, & Workspace |
Watch Video that agents understand |
Scillm 1-shot, exec, subagent, and DAG orch. |
Debugger Live debugger variables for agents |
Sparta Explorer Space-cyber evidence workbench Public product overview for a local-first investigation workspace that traces space-cyber conclusions from framework guidance to program evidence and human review. |
||
| I am looking for... | Go to | What is inside |
|---|---|---|
| A capability to solve a task | skills/ |
Browser automation, memory recall, video analysis, code review, model calls, and other modular capabilities |
| A bounded worker or reviewer | agents/ |
Worker contracts, reviewer roles, stop conditions, receipts, and scheduler definitions |
| Persona patterns and registry | personas/ |
Registry records, memory probes, and voice-readiness evidence |
| Guardrails and session discipline | hooks/ |
Memory-first behavior, quality gates, and completion discipline |
| Maintenance triage leads | reports/agent-maintainer/ |
Report-only surfaces; warnings are leads for human review |
| I want to... | Touch this | Because |
|---|---|---|
| Invoke a capability | skills/<name>/run.sh + skills/<name>/SKILL.md |
Skills own executable behavior, routing rules, artifacts, and proof language |
| Learn how a capability works | skills/<name>/README.md |
Skill READMEs are operator-facing guides |
| Repair or extend a capability | skills/<name>/SKILL.md, scripts, tests, sanity.sh |
Contract changes need focused implementation and proof |
| Delegate bounded work | agents/<name>/AGENTS.md + persona.yaml |
Agents define ownership, denied scope, allowed skills, receipts, and stop conditions |
| Inspect persona availability | personas/registry.yaml + personas/README.md |
Personas are registry records, not generated corpora |
| Enforce session behavior | hooks/ |
Hooks inject memory, block unsafe exits, and enforce evidence discipline |
| Make maintenance decisions | reports/agent-maintainer/latest.md |
Reports are decision surfaces, not dashboards |
Latest sweep: msh-20260628-084135
| Inventory | Count |
|---|---|
| Skills | 347 |
With run.sh |
293 |
With sanity.sh |
292 |
| Agent directories | 92 |
With AGENTS.md |
86 |
With persona.yaml |
51 |
With services.yaml |
7 |
Health check: 171 healthy, 158 warnings, 1 critical. These are triage numbers, not quality scores. If you are looking for a way to learn the repo, the latest report is a useful place to browse possible follow-up work.
Reports live at:
reports/agent-maintainer/latest.md
reports/agent-maintainer/latest.json
reports/agent-maintainer/latest.html
reports/agent-maintainer/runs/<run_id>/
agent-skills/
skills/ reusable capabilities with SKILL.md contracts
agents/ bounded workers, reviewers, monitors, maintainers
personas/ persona registry, schemas, and guide
hooks/ lifecycle gates and prompt hooks
scripts/ repo maintenance, checks, and queue runners
reports/agent-maintainer latest repo sweep for human decisions
deploy.sh broadcast skills and hooks to local agent homes
workers-registry.json generated worker registry
skills/ holds 347 capabilities. Each durable skill is a directory with a
contract, entrypoint, and proof surface:
skills/<name>/
SKILL.md the contract; read this first
README.md human guide when you need context
run.sh the stable entrypoint
sanity.sh cheap local proof
scripts/ implementation details
references/ schemas, examples, templates
Start with the skill list before writing a parallel utility. Browser automation, LLM calls, memory recall, report writing, GitHub tickets, code review, video analysis, and scheduled monitoring already live here, and many skills are easy to adapt.
Typical invocation:
cd skills/surf
./run.sh tab.list --jsonagents/ holds 72 bounded workers. Unlike skills, agents are not broadcast; they
are referenced by project-agent configuration and maintainer routing.
A good agent has a narrow boundary:
- what it owns
- what it does not own
- which skills it may call
- what artifacts prove a turn
- when it must stop or ask for help
Normal shape:
agents/<name>/
AGENTS.md
persona.yaml
services.yaml only for scheduled jobs
Missing persona.yaml is a normalization candidate, not proof the worker is
broken.
personas/ is the roster and contract layer for persona work, not a storage dump
for generated character assets.
personas/
README.md
registry.yaml
schemas/persona-registry.schema.json
The registry points to source directories, memory probes, and voice-readiness
receipts. Runtime memory lives in $memory; generated media and voice artifacts
live under their own storage or skill job paths.
Hooks run around agent lifecycle events. They keep sessions grounded:
- memory first: start with context
- command safety: keep guardrails around risky shell actions
- honest exits: if something failed, say so
- evidence required: completion claims need proof
Deploy:
./deploy.sh --hooksImportant paths:
hooks/memory-first.sh
hooks/quality-gate.sh
hooks/task-complete-gate.sh
hooks/prompts/
Three roles, one issue at a time:
| Role | Owns | Does not own |
|---|---|---|
agent-maintainer |
Scheduled sweeps; latest reports; preview-first ticket drafts | Automatic deprecation, deletion, repair, or issue closure |
monitor-skill-health / monitor-sparta |
Normalized findings, manifests, and $ticket handoffs |
Patching leased targets or closing issues |
agent-skill-maintainer (agents/skill-maintainer) |
One GitHub issue lease at a time; Tau handoff; repair routing; verifier/review/WebGPT evidence bundle; proof-based issue disposition | Broad untracked cleanup, multi-issue repair, closure from WebGPT alone |
Run the report-only sweep:
mkdir -p reports/agent-maintainer
skills/monitor-skill-health/run.sh audit \
--no-memory \
--no-deep-review \
--repo-report \
--json > reports/agent-maintainer/last_run.jsonDraft tickets from the latest concrete monitor violations:
skills/monitor-skill-health/run.sh tickets --jsonCreate those tickets only after reviewing the preview artifact:
skills/monitor-skill-health/run.sh tickets \
--repo grahama1970/agent-skills \
--applyThe scheduled job is registered in agents/agent-maintainer/services.yaml but
disabled by default. Enable it only when the scheduler environment is ready
and you want the cron active.
You do not need the full private runtime to make the repo better. Useful contributions include clearer docs, tighter contracts, better examples, safer scripts, and source-level logic fixes.
When private infrastructure blocks local execution, say that in the PR or issue
and include the checks you could run. For behavior changes, update the relevant
SKILL.md or AGENTS.md in the same change set so the guide and implementation
stay together.
Add a skill when the capability is reusable and needs a contract, entrypoint, and proof. Add an agent when a bounded role, receipts, and stop conditions matter.
One-off scripts, product features, broad cleanup, and standalone prompts belong elsewhere unless they fit an existing skill or agent contract.
- Put the operating contract in
SKILL.md - Use
run.shas the stable entrypoint - Add
sanity.shfor cheap local proof - Reuse existing skills before adding infrastructure
- Keep generated outputs and heavy artifacts out of the skill directory unless the contract explicitly says otherwise
- Define the owner boundary in
AGENTS.md - Encode the role in
persona.yaml - Add
services.yamlonly for scheduled jobs - Avoid agents that own detection, mutation, queueing, and closure all at once
The maintainer sweep was local and deterministic: no mocks, no live calls, and no exercise of runtime behavior.
| What was checked | What was not |
|---|---|
347 skills via monitor-skill-health audit |
Semantic correctness of each skill |
Shallow agents/ metadata inventory |
Runtime behavior of each agent |
| Live scheduler daemon registration | |
| Live GitHub issue mutation |
Rerun the sweep after substantial changes. Use
reports/agent-maintainer/latest.md before final maintenance decisions.
────────
This public repository is for reusable agent patterns, contracts, prompts, hooks, and maintenance workflows. It is intentionally decoupled from controlled, proprietary, ITAR-restricted, or sensitive operational data. Private runtime context, credentials, regulated artifacts, and deployment-specific details stay outside this repo.