A personal collection of original agent skills, engineering workflows, and selected developer tools. One place to explore the work, inspect the source, and share a useful approach from a profile or an article.
Explore the website · Browse categories · Browse the skills · Follow the workflow · Read the engineering article
Explore Roles to select a responsibility, expected handoff and relevant canonical skills. Eight presets cover frontend, Go, architecture, QA/accessibility, security, research, video and release verification. They are not additional agents or permission grants.
Eight new original skills cover codebase orientation, code-index qualification, reference-video analysis, media timeline assembly, continuity review, role composition, dataset readiness and engineering experiment analysis. Code Intelligence, Engineering Experiments and the expanded media bundle connect them to existing procedures.
node tools/bundle.mjs roles
node tools/bundle.mjs role frontend-engineer
python3 examples/specialist-collection/demo.pyRole resolution is read-only. The synthetic demonstration checks actual local source/narration changes and describes a fixed workload without hiding failed tasks. No upstream indexer, renderer, persona library or scientific stack is installed. Contracts and limitations · Role binding.
Explore Frontend Engineering, Modern React Applications, Frontend Performance and Reliability and Interface Design Systems. The worked guide connects state ownership, data failures and accessible icons.
Twenty-five original skills cover React/Next.js, TanStack Query, Tailwind, forms, state/data/navigation boundaries, browser storage/workers, performance/testing/security, and design systems/icons/motion. A separately locked optional React fixture exercises actual React, Query, Hook Form, Tailwind and React Icons; it is not installed by the Astro build or bundle resolver. Coverage and limits distinguish those tests from agent-host or production certification.
112 original skills · 8 specialist roles · 15 primary categories · 18 overlapping bundles · 6 original utilities · 3 selected external tools.
Browse the category directory by the problem you need to solve. Each entry has one primary category, so category totals do not double-count it. Bundles deliberately overlap because the same skill can be useful in several engineering workflows.
Start with TypeScript, Go, contracts and data, durable workflows, parallel engineering, infrastructure, media/provider pipelines, MCP tooling, agent orchestration, or agent runtime and governance.
Each skill is maintained under skills/. Bundles reference canonical files; private adapters supply repository paths, commands, accepted contracts, and policies without forking the public instructions. TypeScript 7 qualification is covered without silently changing this Astro site's compiler dependency.
node tools/bundle.mjs list
node tools/bundle.mjs resolve typescript
node tools/bundle.mjs resolve agent-orchestration --require-cleanThe resolver uses Node built-ins and prints selection/file hashes, including required companion modules. It does not install skills, call models, run selected tools, or start an MCP server. See bundle contracts, private adapter design, category contracts, and coverage and evidence.
The 106 domain, orchestration, source-access, qualification, delivery and frontend additions contain 318 synthetic trigger, boundary, and non-trigger scenario inputs. Those inputs are not completed agent-host runs; instruction effectiveness remains experimental.
Agentflow validates task graphs, matches declared worker capabilities, accounts for resource/token reservations, checks run-record consistency, and proposes reuse only after matching accepted artifact hashes. It does not discover or launch actual workers, execute task check commands, invoke paid models, or authenticate locally supplied approvals.
node tools/agentflow.mjs --help
node tools/agentflow.mjs plan examples/agentflow/plan.json
node tools/agentflow.mjs schedule examples/agentflow/plan.json examples/agentflow/capabilities.json
node examples/agentflow/demo.mjsThe example workers and records are synthetic. The demonstration shows independent work, review after accepted results, and rejection of stale artifacts. Actual worker adapters, atomic resource allocation, credentials, and permission enforcement belong in the consuming private environment. Hash-linked records are not signed attestations or distributed locks.
| Skill | What it produces | Companion check |
|---|---|---|
| Work Packet Planner | A bounded task with ownership, dependencies, and acceptance criteria. | Reject cycles, overlapping parallel ownership, and unmapped criteria; compute dependency waves. |
| Worktree Handoff | A commit-specific account of changes, checks, and unresolved work. | Inspect actual Git state, both sides of renames, and out-of-scope paths without cleanup. |
| Affected Verification | Focused checks tied to actual execution and measured source state. | Record process results and reject stale or invalidated receipts. |
| Evidence First Debugging | A supported diagnosis and bounded repair tied to the original reproduction. | Reject inconsistent fixed conclusions; run controlled asynchronous defect probes. |
| Behavior Test Design | Requirement-linked cases, independent expected results, and prohibited effects. | Detect uncovered requirements and incomplete case contracts. |
| Interface Quality Review | Reproducible rendered-interface findings and scoped verification. | Keep missing keyboard, interaction, layout, or console evidence visible. |
The instructions, helper code, and examples were written for this collection with AI assistance. Helpers have executable fixture tests; agent-host effectiveness remains experimental. A structural checker cannot authenticate evidence or establish that an agent performs better than another skill. Simple tasks can still use a brief natural-language plan.
From a full checkout, use Node >=22.12. Git is required for repository-state and process-receipt commands and strict bundle provenance. No npm installation, model API, or hosted service is needed for these examples:
node tools/skillcheck.mjs --help
node tools/skillcheck.mjs plan examples/skillcheck/plan.json
node examples/skillcheck/defect-demo.mjs
node --test tests/skillcheck.test.mjs tests/bundles.test.mjs tests/agentflow.test.mjs tests/categories.test.mjsThe Skillcheck demo detects three deliberately seeded defects in small original search/lease functions and checks six corrected cases. It is not a benchmark against another repository. The helper test suite uses disposable Git repositories and real child processes.
Helper commands and full contracts · Evaluation and comparison method
The primary catalog contains local instructions and examples. External executable tools remain separately identified with their real authors and sources. The website does not install tools. Skillcheck only executes a command when explicitly invoked with run; it inherits the caller's permissions and is not a sandbox. Agentflow and Bundle Resolver do not execute selected task commands.
Original authoring approach · Catalog metadata · Getting started · Roadmap · Contribution guide
Use Node 24 and the pinned package manager, pnpm@12.4.2.
npm install --global pnpm@12.4.2
pnpm install --frozen-lockfile
pnpm devpnpm validate # Catalog, package, category, bundle, and README checks
pnpm test # Catalog and executable helper tests
pnpm check # Astro / TypeScript checks
pnpm build # Static output in dist/
pnpm exec playwright install chromium
pnpm test:site # Browser tests against the production buildThe site uses /agent-toolkit/ as its base, including during local development. Repository checks and browser tests do not evaluate agent-host behavior.
The deployment workflow builds and publishes changes that reach main. GitHub Pages uses GitHub Actions as its source. No database, hosted MCP server, paid CMS, or Notion credentials are needed by the site. The post-deployment check requires the exact built revision and expected public inventory before reporting publication success.
The collection is maintained for Jordan Davis. Each original package has its own instructions, examples, and evidence requirements. Actual reused materials and dependencies retain applicable notices and terms. See ATTRIBUTION.md, docs/ORIGINAL_SKILLS.md, and AGENTS.md.
A license for this repository's original content has not yet been selected. Public source visibility does not itself grant a reuse license. External projects retain their own terms; none are relicensed by this repository. Never publish private configurations, credentials, company code, or unapproved knowledge-base content here.
The original Sourcekit companion adds bounded public HTTPS reads, RSS/Atom/JSON Feed snapshots and comparisons, and WebVTT/SRT evidence parsing. Browse the bundle. Authenticated platform adapters remain separate; source routing is not a claim of installed access. Python 3.10+ is required for this optional companion and its tests.
Agent-Reach architecture research records the pinned reference and the differences in our original implementation.
Sourcekit 0.2 separates capture success from a conservative content assessment. Explicit --require-content checks declared literal expectations without claiming semantic truth. Optional link references never trigger a crawl. Four original skills cover browser-session isolation, structured extraction, bounded crawl planning and MCP server qualification.
Implementation and limits · Optional pinned Scrapling parser evaluation. The optional parser environment is separate from the website and default tools; it adds no browser or MCP service.
The Harness Engineering bundle combines six original qualification skills with existing context, lifecycle, recovery and evaluation procedures. Harnesskit provides actual Git observations, supplied-profile/readiness comparisons, saved Codex exec event imports, and a programmatic controller test interface. It is not another controller or a live agent launcher.
node tools/harnesskit.mjs --help
node tools/harnesskit.mjs observe --root .
node tools/harnesskit.mjs events examples/harnesskit/exec.jsonl --format codex-exec --binding examples/harnesskit/binding.json
node examples/harnesskit/controller-demo.mjsThe controller demonstration tests an original synthetic fixture and seven deliberate faults, not Symphony or Contrabass. The optional ast-grep recipe is isolated under evals/structural-refactoring/. No default helper installs software or invokes paid inference.
Harnesskit now checks explicitly selected context files (including selected ignored inputs) and reports bounded canaries without hiding unfinished work or missing usage. Contracts and limits · Bounded Harness Canary.
node examples/harnesskit/evidence-demo.mjs
node tools/harnesskit.mjs canary examples/harnesskit/canary.jsonThe example numbers are synthetic, not measured engineering performance. These additions do not start workers or replace a controller.
Six original skills cover date-windowed investigation, debt triage, evidence-backed visual explanations, product storyboards, media output verification and search discoverability. Browse the bundle or inspect the synthetic worked example.
Publicationcheck checks local evidence records and static HTML, and optionally probes/decodes short local media. It does not search, render, install packages, launch agents or publish. Semantic support, rights, runtime adoption and search ranking are not inferred from mechanical checks.