A curated set of agent skills I actually use, published in the portable
Agent Skills format (a folder per skill with a
SKILL.md). They are written to be workspace-agnostic: no project law, product
names, or private tooling baked in.
The theme is rigor with calm: make "looks done" stop passing for "done", make "looks plausible" stop passing for "true", keep long work moving, and keep the language between humans and agents clear enough that neither side pressures the other.
Register this repository as a plugin marketplace, then install:
/plugin marketplace add aewing/skills
/plugin install rigor-skills@aewing-skills
or from the shell:
claude plugin marketplace add aewing/skills
claude plugin install rigor-skills@aewing-skills
git clone https://github.com/aewing/skills.git && cd skills && ./install.sh
Installs to ~/.claude/skills by default; point it anywhere for other
harnesses:
./install.sh --target ~/.config/your-agent/skills
./install.sh --list
./install.sh --dry-run
| Skill | What it is for |
|---|---|
| execution-rigor | Long-horizon execution that finishes for real: observable definition-of-done, per-task verification, shortcut tripwire scans, and an adversarial pre-completion audit. Use when the failure mode is "called done but actually half-finished." |
| truth-loop | Reasoning that converges on truth rather than plausibility: hypotheses kept alive, evidence that can hurt, contradiction ledgers, and a falsifier on every conclusion. Use when the failure mode is "looked defensible, but reasoned around the hard part." |
| goodtalk | Audit and recode instruction language so humans and agents stop accidentally pressuring each other ("don't fuck with the model"). Research / standards / suggest / execute / recode / install / sync / publish modes with a portable standards contract. |
The pair works together: execution-rigor guards the doing, truth-loop
guards the thinking. goodtalk keeps the instructions they run under honest
and calm.
| Skill | What it is for |
|---|---|
| keep-going | Maintain progress on multi-step, long-running work: explicit remaining state, one causal unit at a time, milestone reviews, honest handoffs. |
| coherence-engine | Focus and pruning for long-horizon tasks: flow perception, cut what does not serve the core intent, autonomy with a clear reversible/irreversible line. |
| verification-hygiene | Prove claims against the real artifact: the narrowest falsifying check, required consequences per change type, and rejection of false green. |
| adversarial-reviewer | Evidence-first review with one verdict (SHIP / BLOCKED / REWORK / MORE PROOF), highest blast radius first, smallest honest remediation. |
| focus-group | Run clearly labeled synthetic user panels with causal participant selection, complete bounded transcripts, evidence-aware findings, and predictable project-owned outcomes. |
| Skill | What it is for |
|---|---|
| wat | ADHD-friendly response overlay: answer first, one screen, evidence kept, one next action. For status, summaries, and handoffs. |
| roast-review | Evidence-backed critique: funny, ruthless, specific reduction to the target's weaknesses and self-owns, plus a concrete escape plan (ROAST.md / PROPOSAL.md / EVIDENCE.md). |
| tui-design | Terminal UI craft: layout paradigms, keyboard and focus models, semantic color, animation, and anti-pattern and compatibility checklists. Framework-agnostic. |
| headless-lanes | Run a mission through headless agent CLIs (any providers) in parallel lanes from one overseer session: recon and callable-model proofs, scorecard routing by the user's quality/speed/cost order, rooms and lane briefs, fixed signals over the project's coordination substrate, dead-air monitoring, verified review and landing. |
Each skill is a plain folder; the source of truth is the SKILL.md plus any
references/ it links. Improvements usually land as: a sharper description
(that is what triggers the skill), one more "when NOT to use this" line, or a
smaller replacement for a wordy rule.
The repo governs itself: the active language profile is
.goodtalk/standards.json, and goodtalk runs against this repo use it. To set
up the same for your project, copy the starter at
skills/goodtalk/examples/standards.json.starter or run goodtalk standards init.
Marketplace users:
/plugin marketplace update aewing-skills
Copy-install users: git pull in the checkout, then run ./install.sh again.
MIT — the skills are instruction and process text; take them, adapt them, and publish what you learn. See LICENSE.