Pratham Dubey's skills

A small set of coding-agent workflows for planning, git work, code review, delegation, and session continuity. Each skill has a narrow job, explicit stop conditions, and defaults listed in its own file.

Install

skills.sh

npx skills@latest add prathamdby/skills

Pick the skills and agents you want in the installer.

Claude Code

/plugin marketplace add prathamdby/skills
/plugin install skills@pratham-skills

Codex

codex plugin marketplace add prathamdby/skills
codex plugin add skills@pratham-skills

Quickstart

• /prath-mode routes work to one skill or a delivery chain.
• /peer-review checks an implementation plan before work starts.
• /verify fans out N isolated attempts, ranks them, and stops on a named gate.
• /deslop removes needless complexity from a selected diff.
• /commit creates a clean-room commit from staged or tracked unstaged work.
• /make-pr pushes committed work and creates or updates its pull request.
• /fix-pr hunts, triages, fixes, and replies to pull-request feedback and CI.
• /gh orients on a PR, review threads, or red CI, or posts one reply.
• /explain-diff writes a self-contained HTML walkthrough of a change.
• /recon maps the current repository and refreshes only changed areas later.
• /box clones and searches an external git repository locally.
• /assign runs one exact task through an external coding-agent CLI.
• /handoff saves resumable session state or continues from it.
• /orchestrate coordinates cheaper subagents while the main agent verifies.

Why these skills exist

Common failure	Skill	Contract
The agent picks the wrong workflow or repeats work owned by another skill.	prath-mode	Routes each immediate action to one owner and tracks chain completion.
A plan misses a requirement or carries a risky assumption into implementation.	peer-review	Exhausts every material finding, ranks them, and issues a fixed verdict. It edits only with explicit authority.
A plan review stops after the first risk and leaves other blockers unlisted.	peer-review	Surfaces every material finding with no count cap, then maps the full ranked list to the verdict.
The agent ships the first attempt or picks a winner from narration instead of observed evidence.	verify	Fans out N isolated attempts, pairwise-verifies on decomposed criteria, ranks with PPT, and stops on a named gate.
Generated code adds guards, wrappers, comments, or indirection that the codebase does not need.	deslop	Classifies the selected diff against six categories, preserves staging intent, and verifies behavior-sensitive edits.
Commit messages leak ticket or review context and do not match the committed hunks.	commit	Locks the snapshot, traces every message line to a hunk, applies the selected hook policy, and verifies the commit.
PR creation misses local commits, duplicates an existing PR, or describes work absent from the diff.	make-pr	Blocks on a dirty or diverged branch, publishes committed work, reuses the open PR, and verifies its fields.
A PR body stays flat for large diffs or buries small changes in boilerplate.	make-pr	Measures the locked diff, picks a body tier for bullet count and depth, and writes STE100 and Google-style prose.
A PR body of any tier leaves a shape change as prose when a sketch would make the delta clear.	make-pr	Drafts rich proved views from Body visuals, including on S, and places each view next to the theme it supports.
A harness appends marketing footers such as Made with Cursor to the PR body.	make-pr	Requires the published body to equal the ledger body and strips injected footers once before success.
Review work starts from the first visible comment and misses later pages, nested replies, or invalid suggestions.	fix-pr	Exhausts every feedback surface before editing, requires evidence for each verdict, and re-hunts until stable.
Required CI or check runs on the PR head are left red while only human comments are fixed.	fix-pr	Hunts terminal non-success required or blocking checks and annotations with the other surfaces, then triages and fixes them.
A fix-pr commit subject narrates review follow-up instead of the locked hunks.	fix-pr	Discards pre-drafted subjects, requires a hunk-proved subject recipe, and blocks ban-list or conversation-only messages before push.
A fix-pr commit picks up identity or harness trailers from hooks or agent defaults.	fix-pr	Denies Co-authored-by / Signed-off-by / Made-with and freeform harness footers by default and runs commit Trailer hygiene before push.
Agents burn 3–5 gh calls and dump CI logs into context when inspecting a PR.	gh	One script per I/O loop; bounded snippets; logs on disk; raw gh only after gotchas.
A fix-pr hunt runs gh scripts without loading the gh skill, then invents raw gh to reply.	fix-pr	Requires the gh skill before any GitHub I/O and posts replies through pr-reply.ts.
Agents skip the .ts inspect scripts on Node 22 and dump GraphQL instead of probing bun, nub, tsx, or type-stripping Node.	gh	scripts/run tries bun, nub, tsx, then Node (native TS or --experimental-strip-types, including nvm); GraphQL inspect is blocked until that list is exhausted.
A large diff gets a shallow chat summary with no surrounding system context.	explain-diff	Groups the change by theme and writes an evidence-linked HTML page with a working quiz.
Every session re-reads the same repository from scratch.	recon	Stores a bounded evidence map and patches it from committed git drift.
The agent guesses what an external repository contains.	box	Clones into a skill-owned sandbox, searches local source, and returns cited findings.
External coding-agent commands break on quoting, permissions, silence, or parallel runs.	assign	Uses collision-safe stdin transport, non-interactive commands, tracked processes, cleanup, and result verification.
A resumed session trusts stale paths, tasks, branches, or PR state.	handoff	Saves a bounded, redacted handoff and validates every artifact before resuming work.
The main model spends its context on mechanical work or trusts delegate summaries.	orchestrate	Delegates disjoint chunks, verifies evidence and integration, and keeps the parent read-only.
Best-of-N collapses to the first plausible attempt or one yes/no judge call.	verify	Fans out a fixed-N pool, scores pairs on a 20-letter scale, ranks with a pivot tournament, and stops on a named gate.

Reference

Skill	Description	Flags and arguments
prath-mode	Route one action or a complete workflow chain.	None
peer-review	Exhaustively review a plan or proposed change and issue a fixed verdict.	None
verify	Fan out N isolated attempts, pairwise-verify, and select, or score current progress.	--candidates default 3, --evals default 2, --pivots default 2, --max-rounds default 0, --criteria default auto, --track
deslop	Remove code slop from one git diff without changing behavior.	--staged default, --unstaged, --base <branch>
commit	Commit a locked snapshot with hunk-traced copy.	--staged default, --unstaged, --conventional default, --simple, --verify, --allow-trailers
make-pr	Publish a branch and create or update its PR with a diff-scaled body and rich proved visuals on every tier.	--target <branch> default main, --ticket <id>, --conventional
fix-pr	Resolve open PR feedback and CI, then reply with evidence.	--pr <n|url>, --no-push, --no-reply
gh	Orient on PR state, review threads, or CI, or post one reply.	--json, --full, -R owner/repo, threads --open/--all/--complete, CI --pr/--sha/--list, reply --in-reply-to/--conversation/--body-file/--body; invoke via scripts/run
explain-diff	Write an HTML teaching page for a diff, branch, or PR.	--target <branch> default main, --pr <n|url>, --staged, --unstaged, --output <path>
recon	Build or refresh a persistent map of the current repo.	--refresh, positional focus
box	Clone, update, list, search, or persist an external repo.	--persist, --update, --list, --no-subagents
assign	Run one task with OpenCode, Codex, or Claude Code.	--agent <name> default opencode, --model <model>, --dir <path>
handoff	Save or resume bounded session state.	--resume <path>, --path <path>, positional focus
orchestrate	Coordinate in-harness subagents as a read-only parent.	Positional task

Development

Before committing a skill edit, run the manual checks in AGENTS.md:

1. Frontmatter name matches the skill directory.
2. Description is present and at most 1,024 characters.
3. SKILL.md is at most 100 lines.
4. Quickstart and Reference both include the skill.
5. Every real Markdown path in the skill resolves.

License

MIT © 2026 Pratham Dubey