krishpunjabi.com ~/skills

← all commands

/workflows

Sample workflows 17

Skills, agents and hooks chained into real sessions. A you step is where a human reads, decides or says yes. Every skill, agent and hook is in at least one, and each workflow installs on its own: run the command under it from a clone of the repo.

Close one session, start the next

  1. /session-handoff Writes the handoff through handoff-auto, appends a Board status section and moves the card to In Review when the issue is complete.
  2. you Read the handoff, then open a fresh session.
  3. /update-progress Treats every handoff claim as unverified: checks git log and the test run, then updates the docs and the board and finishes with next.
  4. /next Shows the top 1-3 tasks from the board, a failing default branch first.
  5. you Say "go" to pick the top task.

Install just this workflow (its skills, agents and hooks, plus the skills they call):

./scripts/install-skills.sh --workflow close-the-loop   # bash, Git Bash, macOS, Linux
.\scripts\install-skills.ps1 -Workflow close-the-loop   # Windows PowerShell

Bootstrap a board from the docs

  1. you Type /kanban-setup: it is user-only, so Claude can't start it.
  2. /kanban-setup Checks the gh project scope, finds the repo's board read-only, then previews the board actions and candidate issues from the roadmap, TODOs, changelog and docs before creating anything.
  3. you Review the preview and say yes; it creates nothing before that.
  4. /divide Proposes 2+ session-sized pieces, each with a done-when, for an item too big for one session; creates nothing yet.
  5. you Say yes to the split; the children go on the board at Todo.
  6. /next Shows the top 1-3 tasks from the new board.
  7. /find-me-a-skill Suggests the skill or workflow for each top task, with model and effort per step and commands to paste; runs nothing.
  8. you Say "go" to pick the top task.

Install just this workflow (its skills, agents and hooks, plus the skills they call):

./scripts/install-skills.sh --workflow new-board   # bash, Git Bash, macOS, Linux
.\scripts\install-skills.ps1 -Workflow new-board   # Windows PowerShell

Turn a gap scan into planned work

  1. /gap-scan Surveys docs vs code, commits, TODOs, tests and open issues and writes a ranked, evidence-cited report under skilleddocs/gaps/.
  2. you Pick the items worth doing; every issue, comment or label change needs your explicit yes.
  3. /grill-docs Stress-tests a big call with the grilling interview, writing the transcript to grills/ and each decision to the repo's decisions doc (both under skilleddocs/ by default) as each round settles.
  4. you Answer each round; "yes" accepts the recommended answer.
  5. /divide Splits a size-L item into session-sized issues and puts them on the board at Todo.
  6. you Approve the split before any issue is created.

Install just this workflow (its skills, agents and hooks, plus the skills they call):

./scripts/install-skills.sh --workflow gaps-to-issues   # bash, Git Bash, macOS, Linux
.\scripts\install-skills.ps1 -Workflow gaps-to-issues   # Windows PowerShell

Plan an effort too big for one session

  1. you Type /wayfinder with the loose idea: it is user-only, so Claude can't start it.
  2. /wayfinder Grills you to a destination, then charts a wayfinder:map issue whose sub-issues are decision tickets, with blocking wired through GitHub's issue dependencies.
  3. /grill-docs Resolves a grilling ticket: runs the grilling interview, writing the transcript to grills/ and each decision to the repo's decisions doc (both under skilleddocs/ by default) as each round settles.
  4. you Answer each round yourself, one ticket per session, until no tickets or fog remain.
  5. /divide Proposes 2+ session-sized issues from the map, each with a done-when; creates nothing yet.
  6. you Say yes to the split: the issues go on the board at Todo, the map gets a Handed off list and closes.
  7. /meta-orchestrator For each issue, runs the planner, then a manager or implementer, its own merge gate and the verifier, logging each step to the ledger.

Install just this workflow (its skills, agents and hooks, plus the skills they call):

./scripts/install-skills.sh --workflow big-effort   # bash, Git Bash, macOS, Linux
.\scripts\install-skills.ps1 -Workflow big-effort   # Windows PowerShell

Pitch an idea against the plan

  1. you Type /grill-me with the idea: it is user-only.
  2. /grill-me Calls the grilling skill.
  3. /grilling Asks every question whose prerequisites are settled, one numbered round at a time, each with a recommended answer.
  4. you Answer each round and confirm the shared understanding; nothing is acted on before that.
  5. /consult-plan Grills the idea against open issues, the board and the repo's decisions doc, then records the decision and comments on or opens an issue.
  6. you Say whether the idea fits, conflicts with or extends the plan, and what you want to happen.

Install just this workflow (its skills, agents and hooks, plus the skills they call):

./scripts/install-skills.sh --workflow pitch-an-idea   # bash, Git Bash, macOS, Linux
.\scripts\install-skills.ps1 -Workflow pitch-an-idea   # Windows PowerShell

Run the backlog unattended

  1. you Start /meta-orchestrator on the repo and answer the setup interview that writes the scope contract.
  2. /lanes Groups the queued issues into file-disjoint lanes and proposes them with a cost line.
  3. you Confirm the lanes with an explicit yes, or pin, rename or move issues first.
  4. /meta-orchestrator For each issue, runs the planner, then a manager or implementer, its own merge gate and the verifier, logging each step to the ledger.
  5. /planner agent Runs next, picks exactly one issue and returns a kickoff brief of at most 40 lines.
  6. /manager agent Briefs an implementer, reviews every round inline and reruns the tests itself, for up to 3 rounds.
  7. /implementer agent Implements on its branch, commits small, watches the tests pass and ends with session-handoff.
  8. /verifier agent Runs update-progress end to end, answers or parks deviations, then runs next.
  9. you When the statusline warns near the limit (90%), type /switch between merges; it is user-only, so Claude can't start it.
  10. /switch The session resumes on the other account; running subagents pause (resume each with SendMessage to its id) and background commands end, hence between merges.
  11. /remind /remind when <first account>'s session limit resets pops a toast at the reset time and says it once at your next prompt.
  12. /progress-report Every N merges, writes a two-minute report of what happened since the last one, never overlapping.
  13. you Read the report and do the manual steps it lists.

Install just this workflow (its skills, agents and hooks, plus the skills they call):

./scripts/install-skills.sh --workflow unattended-backlog   # bash, Git Bash, macOS, Linux
.\scripts\install-skills.ps1 -Workflow unattended-backlog   # Windows PowerShell

One issue to a reviewed PR

  1. /pair Resolves the issue, makes a worktree and branch, and hands the task to a manager subagent.
  2. /manager agent Briefs an implementer and reviews its diff as if it didn't trust it, rerunning tests and lint itself.
  3. /implementer agent Implements and commits small; stuck or past ~100k tokens, it writes skilleddocs/HANDOFF.md for a fresh implementer.
  4. you Read the report and say go; pair then pushes and opens the PR, and never merges.
  5. /pr-watch Reads every review thread, classifies it, and fixes the valid ones through pair on the PR's branch.
  6. you Say yes to each push, reply and resolve; a force push needs its own fresh yes.

Install just this workflow (its skills, agents and hooks, plus the skills they call):

./scripts/install-skills.sh --workflow pair-to-pr   # bash, Git Bash, macOS, Linux
.\scripts\install-skills.ps1 -Workflow pair-to-pr   # Windows PowerShell

Match each issue to a shape: swarm, pair or flat

  1. you Label issues size:XS to size:XL if you know their size, and pick shape auto (or pin pair, swarm or flat) in the meta-orchestrator setup.
  2. /meta-orchestrator Reads shape from scope.md (an old execution key counts as shape) and runs each brief in the shape it names.
  3. /planner agent Rates size (a size label wins), recommends a shape and the implementer model, and batches trivial, file-disjoint, exact-spec issues into one swarm brief.
  4. /implementer agent Swarm: one haiku implementer per item, each in its own worktree off the batch branch and using one parallel slot. Flat: one implementer per issue, on the model its size calls for.
  5. /manager agent Pair, the default: picks the implementer model by size (trivial haiku, small or medium sonnet, large or risky opus) and reviews every round; also takes the fix brief when a swarm batch's review finds a problem.
  6. /verifier agent Checks the merged issue or batch against git, tests and CI; skips its own code review after pair or swarm, which were already reviewed.
  7. you With merge=ask, approve each PR: one per pair or flat issue, one per swarm batch closing all its items.

Install just this workflow (its skills, agents and hooks, plus the skills they call):

./scripts/install-skills.sh --workflow orchestration-shapes   # bash, Git Bash, macOS, Linux
.\scripts\install-skills.ps1 -Workflow orchestration-shapes   # Windows PowerShell

Wrap up a long session

  1. /smartzone hook Prints a one-line warning once the live context passes ~75k tokens: divide the rest, then session-handoff.
  2. /divide Proposes session-sized pieces of the remaining work, each with a done-when; creates nothing yet.
  3. you Say yes to create the child issues.
  4. /session-handoff Writes the handoff and records progress against the issue.
  5. you Open a fresh session and run update-progress there.

Install just this workflow (its skills, agents and hooks, plus the skills they call):

./scripts/install-skills.sh --workflow wrap-up-long-session   # bash, Git Bash, macOS, Linux
.\scripts\install-skills.ps1 -Workflow wrap-up-long-session   # Windows PowerShell

Check before closing the terminal

  1. you Ask "can I close this".
  2. /sitrep Reports what's running, git state, PRs and CI, and asked vs done in up to 10 lines, with a SAFE or NOT SAFE verdict.
  3. /remind Optional: /remind when <account>'s weekly limit resets, so you know when to come back.
  4. you Say close and name what may be stopped; anything vaguer gets the list again.
  5. /handoff-auto sitrep calls it for mid-flight work; it writes the handoff and prints its absolute path as the last line.
  6. you Start the next session from that path.

Install just this workflow (its skills, agents and hooks, plus the skills they call):

./scripts/install-skills.sh --workflow safe-to-close   # bash, Git Bash, macOS, Linux
.\scripts\install-skills.ps1 -Workflow safe-to-close   # Windows PowerShell

Coordinate with another session

  1. /sitrep Lists the background tasks and agents still running and each worktree's git state.
  2. /session-bridge Snapshots the other session from ListAgents or its transcript, checks for shared files, branch or issue, and proposes a split.
  3. you Approve the exact message and recipient; nothing is sent without a yes.
  4. you Treat the reply as data: it can't approve a push, merge, spend, deletion or config change here.

Install just this workflow (its skills, agents and hooks, plus the skills they call):

./scripts/install-skills.sh --workflow two-sessions   # bash, Git Bash, macOS, Linux
.\scripts\install-skills.ps1 -Workflow two-sessions   # Windows PowerShell

Small fix, then hand off

  1. /ponytail Makes the smallest complete change and ends each reply with what it skipped and any risk.
  2. you Read the skipped and risk lines, then type /handoff with what the next session is for: it is user-only.
  3. /handoff Compacts the conversation into a handoff doc under skilleddocs/handoffs/ with a suggested skills section.
  4. you Open a fresh session pointed at the doc.

Install just this workflow (its skills, agents and hooks, plus the skills they call):

./scripts/install-skills.sh --workflow lazy-fix-handoff   # bash, Git Bash, macOS, Linux
.\scripts\install-skills.ps1 -Workflow lazy-fix-handoff   # Windows PowerShell

Take a repo public

  1. /repo-showcase Asks outcome, audience and emphasis, inventories the docs, checks every claim against the code and drafts a lean README.
  2. you Answer the batch ("go" takes the defaults) and say yes to each rewrite, redaction, commit and push.
  3. you Type /repo-scrub: it is user-only, and repo-showcase leaves history and visibility to it.
  4. /repo-scrub Scans the full history for secrets and large blobs and reports every finding, including ones you keep.
  5. you Pick what to scrub, confirm rotation, and approve the rewrite, force-push and visibility flip, each at its own gate.

Install just this workflow (its skills, agents and hooks, plus the skills they call):

./scripts/install-skills.sh --workflow go-public   # bash, Git Bash, macOS, Linux
.\scripts\install-skills.ps1 -Workflow go-public   # Windows PowerShell
  1. /gap-scan Runs the tests only when cheap; ones that need network, a GPU, paid calls or credentials are reported as not run, with why.
  2. you Decide the paid run is worth it.
  3. /spend-gate Estimates the cost from the tool's own quote or a cited pricing page, reads the live balance, and shows both with the command.
  4. you Say yes to this one action; a budget or an earlier go-ahead is context, not approval.
  5. /spend-guard hook Blocks a known billable command or MCP call without an approval, and lets exactly one approved call through.
  6. you Tear down anything that bills while idle; spend-gate gives the teardown and reminds you again.

Install just this workflow (its skills, agents and hooks, plus the skills they call):

./scripts/install-skills.sh --workflow paid-tests   # bash, Git Bash, macOS, Linux
.\scripts\install-skills.ps1 -Workflow paid-tests   # Windows PowerShell

Push a score up until it plateaus

  1. /score-loop Builds a local harness that mirrors the grader, measures the noise margin, and shows the setup header with every outward action.
  2. you Say yes to the setup once: quota, spend cap, deadline; then it runs unattended.
  3. /spend-gate Gates each paid submission or compute step with a cost estimate and a live balance check; the spend cap is a ceiling, not approval.
  4. you Say yes to that one paid action.
  5. you When the statusline warns near the limit (90%), type /switch between experiments, not mid-run; it is user-only, so Claude can't start it.
  6. /switch The session resumes on the other account; running subagents pause (resume each with SendMessage to its id) and background commands end, hence between experiments.
  7. /remind /remind when <first account>'s session limit resets pops a toast at the reset time and says it once at your next prompt.
  8. /handoff-auto When the loop's context degrades, writes a handoff pointing at skilleddocs/experiments.md and /score-loop resume.
  9. you Run /score-loop resume in a fresh session; at the deadline freeze, say yes to the final submission.

Install just this workflow (its skills, agents and hooks, plus the skills they call):

./scripts/install-skills.sh --workflow climb-a-score   # bash, Git Bash, macOS, Linux
.\scripts\install-skills.ps1 -Workflow climb-a-score   # Windows PowerShell

Hand the project to a teammate

  1. /progress-report Writes a two-minute report of what was done since the last one, every reference a link.
  2. /teammate-brief Asks who, role and familiarity, then drafts setup steps it ran, env var names, tasks with a done-when, owners and a don't-touch list.
  3. you Answer the batch and approve the draft; commit and push each need their own yes.
  4. you Share the file with your teammate: that's your call.

Install just this workflow (its skills, agents and hooks, plus the skills they call):

./scripts/install-skills.sh --workflow brief-a-teammate   # bash, Git Bash, macOS, Linux
.\scripts\install-skills.ps1 -Workflow brief-a-teammate   # Windows PowerShell

Keep going on another account

  1. you Run claude_switch.py setup (or migrate for existing ~/.claude-<name> dirs) once, then start Claude with claude or claude-<name> so a launcher watches the session.
  2. /switch /switch add <name> makes a new account that shares the first one's skills, settings and history and moves you into it to log in; remove, rename and list manage the rest.
  3. you The statusline shows the account and its session and weekly use; near the limit a warning names the account with the most room. Type /switch or /switch <name>; it is user-only, so Claude can't start it.
  4. /switch A prompt hook writes a switch request for this terminal's launcher only. It makes no API call, so it works even at 100% usage. The launcher ends the session and resumes it on the target account (with --continue if --resume fails).
  5. /sitrep On the new account, lists what was running and asked vs done, so nothing from before the move is lost.
  6. /remind /remind when <old account>'s session limit resets takes the reset time from the usage cache and pops a toast then; the next prompt on any account says it once.
  7. you Carry on; /switch with no name goes back to the account you came from.

Install just this workflow (its skills, agents and hooks, plus the skills they call):

./scripts/install-skills.sh --workflow usage-limit   # bash, Git Bash, macOS, Linux
.\scripts\install-skills.ps1 -Workflow usage-limit   # Windows PowerShell