atreya2011/agent-skills

Skills for running and tracking coding-agent sessions in herdr, kept free of machine values so that a clone works on any machine. CONTEXT.md is the glossary every skill, brief and issue uses; decisions are recorded in docs/adr/.
Skills
- cf-page-store: Run one private Cloudflare store for phone-page pages and their latest submissions.
- cto: Run the CTO in one herdr tab: rebuild the picture of every session, todo and pull request from sources, close todos on proof, put judgment calls to the user in one batch, and write a brief. A Go snapshot tool gathers the facts.
- find-unknowns: Map unknown unknowns before work by exploring the codebase and domain in parallel, teaching the findings, and writing a persistent map under docs/unknowns.
- gauntlet: Take one approved issue from facts through blind adversarial review to a merge-ready pull request, with its seats read from a seat table in the local file.
- gauntlet-deliver: Deliver a converged gauntlet branch as a pull request marked ready for review, with the review notes of the closing cycle.
- gauntlet-implement: Stage and implement one approved issue into committed, cap-checked code on its own branch, through the implementer tab or parallel subagents in git worktrees.
- gauntlet-review: Run blind review cycles against a locked HEAD with executable checks and review points until a cycle has no major bug.
- herdr-chat: Talk to another agent through its herdr tab, by chat only: find the tab, send one short message, read the reply.
- herdr-orchestrator: Survey all Claude and Codex tabs across herdr workspaces, report activity and blockers, locate tabs, or resume dead sessions.
- major-refactor: Run an explicit managed program for a large or irreversible codebase change, with audits, a committed manifest, risk-ordered waves, subagents, and approval gates.
- no-slop-writing: Draft or rewrite public prose, or edit existing prose to remove AI tells, using the author's actual writing and current intent as the voice reference.
- phone-page: Publish designed read pages and tappable question pages in an interactive session, only for an item that needs the user's decision or an explanation the user explicitly asked for. Never use it for status summaries or job reports, in headless runs (codex exec, claude -p), or in agent-dispatched or subagent runs.
- thermo-nuclear-code-quality-review: Run an extremely strict maintainability review for abstraction quality, giant files, and spaghetti-condition growth.
- where-are-we: Reconstruct how the current session started and where it stands now by parsing its own transcript, then report it as a short readable narrative (ASD-STE100 mechanics, blunt outcome-first voice).
Writing a skill
- SKILL.md holds only the steps and rules every run needs, in at most 500 words.
- Formats, examples, contracts and edge cases go in files under the skill's references/ folder, each linked from the step that needs it.
Local files
A skill that needs machine values (see CONTEXT.md) reads them from ~/.agents/local/ .md, a file this repo never tracks. Such a skill opens with a ## Local file section that names the path, says what the file holds, and tells the agent to read it when it exists. To set up a new machine, create each file named by such a section. Skills with a local file: cf-page-store, cto, gauntlet, herdr-orchestrator, no-slop-writing, phone-page. gauntlet reads its seat table from gauntlet.md, or from gauntlet- .md when a run names an alternate table.