Connor Griffin's agent skills

A portable skill pack for coding agents. Skills live under skills/ and work with clients that support the Agent Skills format. Some skills compose: the UI workflow uses the bundled browser driver for rendered evidence.

Included skills

Skill Purpose Extra requirement
ui-craft Full surface lifecycle: lock a visual spec, build to it, critique with personas, audit against the lock, polish, re-settle drive-local-webapp for rendered review; parallel-agent support is recommended
drive-local-webapp Drive and screenshot a local web app with headless Chromium Node.js 20+; installs Playwright locally
cbm-onboard Index a repository with codebase-memory-mcp and keep it current codebase-memory-mcp on PATH
spin-worktree Create isolated Git worktrees for issue and PR work Git; GitHub CLI only for --pr discovery
scope Triage front door for work that isn't ready to build: classify the dominant uncertainty and route to one specialist skill, including a bundled interview mode —
plan-review Adversarially review a plan or work order with cold agents before building Parallel-agent support recommended; persona-review optional for load-bearing plans; the round-count evidence behind its rules is in docs/review-round-mining.md
persona-review Convene a panel of persistent reviewer personas to review a document and synthesize a verdict Private data repo for persona memory; GitHub CLI for mining real-colleague profiles; parallel-agent support recommended
orchestrate Flip the session into coordinator mode: delegate all real work to sub-agents routed by an empirically benchmarked model capability table Codex CLI for GPT-tier sub-agents; table re-benchmarked per its bundled procedure
wayfinder Chart a large, foggy effort as a GitHub map of decision tickets, then hand clear subtrees to implementation GitHub CLI; composes with research, scope's interview mode, ui-craft
pr-body Write a PR body and score it before the PR opens, or audit an existing one, against a deterministic linter and a voice judge A PreToolUse hook that hard-denies gh pr create until the body is scored, installed separately from hooks/
research Investigate a question against primary sources and capture findings as Markdown in the repo —
prototype Build a throwaway prototype — a terminal app for logic questions or switchable UI variants in the real app —
tdd Test-driven development through public interfaces, red-green-refactor —
review Review changed code against the project's standards and the originating issue —
implement Implement a PRD or issue set via tdd, then review, then commit tdd and review from this pack
say-less Answer-first output shaping with a per-repo glossary of approved terms; adapted from i-have-adhd, merged with this pack's reader-preference rules and an ASD-STE100 subset Optional prompt hook for always-on digest; see docs/overlay.md and the measured comparison
code-review Review changes since a fixed point on two axes, Standards and Spec, each returning a verdict per enumerated item so a round terminates Parallel-agent support recommended; GitHub CLI to fetch an originating issue; see docs/review-round-mining.md for the mined evidence behind the fix protocol and the round cap
writing-for-agents Write or review a skill against a shared pointer/load/pruning vocabulary; doctor verb audits the loaded skill and CLAUDE.md/AGENTS.md estate for token cost and trigger risk Hand-only; invoke by name

Install

Install the primary UI workflow with its required browser driver using the standard skills CLI:

npx skills add ConnorGriffin/skills \
  --skill ui-craft \
  --skill drive-local-webapp

Install another skill by itself:

npx skills add ConnorGriffin/skills --skill spin-worktree

Install every skill:

npx skills add ConnorGriffin/skills --all

The CLI prompts for the target agents and whether to install per-project or globally. Review skill instructions and scripts before running them; several skills intentionally start processes, modify repositories, or install Git hooks.

Optional upstream skills

Optional integrations are enhancements, not hidden runtime requirements:

  • Codebase Memory: accelerates structural exploration. Without it, use ordinary repository search and file reads.
  • codebase-design: referenced by tdd; helps shape non-trivial module interfaces. Without it, keep the shipping module shape and preserve locality.
  • domain-modeling: referenced by scope's interview mode and wayfinder; maintains a project's domain vocabulary. Without it, ground terms in the repo's own docs.
  • impeccable: adds a final visual-quality audit. Without it, run the explicit contrast, focus, overflow, and target-size checks.
  • persona-review: referenced by plan-review for load-bearing plans; convenes a panel of persistent reviewer personas. Without it, plan-review proceeds on its own fresh-cold-pass termination alone.

Install Matt Pocock's optional skills from their repository:

npx skills add mattpocock/skills --skill codebase-design --skill domain-modeling

Attribution

The implement, tdd, review, and prototype skills are adopted from — and scope's interview mode and code-review are derived from — Matt Pocock's skills repository (MIT, copyright (c) 2026 Matt Pocock), lightly edited here to be self-contained. See LICENSE and NOTICE.

Output style

output-styles/say-less.md is a Claude Code output style: it replaces the built-in coding instructions so every reply leads with the outcome and stops there. The say-less skill is the full ruleset for documents; the style is the always-on subset.

Same task, same model, style the only difference. Neither reply is truncated:

Two replies to one diagnosis task, side by side. The default style answers in 235 words across four sections; say-less answers in 56 words. Both name the off-by-one loop, the loose assertion and the wasted backoff sleep.

Both arms found the same three defects. Full text of every arm: examples/transcripts/.

Across 120 headless runs on a 10-question set with an executed answer key:

model median output tokens, default say-less outcomes correct, default say-less
claude-opus-5 348 164 20/20 10/10
claude-sonnet-5 436 164 115/120 38/40
claude-fable-5 456 178 20/20 10/10

Half to two thirds of the reading, no correctness cost on opus or fable. Sonnet keeps one residual failure worth knowing about before you install it, and compression drops caveats a longer reply would have kept: both are measured, with their method and the counterexamples, in docs/before-after.md.

Activate it by setting "outputStyle": "say-less" in ~/.claude/settings.json after linking the file into ~/.claude/output-styles/, or run /output-style say-less in an interactive session. New sessions pick it up; the one you are in does not.

Shared profile

profile/ holds a machine-agnostic global-instructions layer (communication rules, working preferences, an engineering charter) meant to be @-imported from a local checkout of this repo into each machine's own CLAUDE.md/AGENTS.md, with machine-specific overlays staying private. See profile/README.md for setup.

Development

Run the repository checks with:

python3 scripts/validate.py

Contributions require a Signed-off-by trailer under the Developer Certificate of Origin.

License and support

MIT, from v0.2.0 onward (v0.1.0 was published under Apache-2.0, and that grant stands for copies taken at that tag). See LICENSE and NOTICE. Security reports belong in the private channel described in SECURITY.md; everything else uses GitHub issues.