agentleFS
seankoji-comseankoji-com/.agents/skills/backlog-burn/SKILL.md

backlog-burn

Skill2 starsChanged 3 days agoMITView on GitHub

Burn a repository's open issues and pull requests down to zero, or launch one run per repository for a named organisation. Use for backlog reduction, clearing open PRs, unblocking owner decisions or dispatching issue fixes. Prefer evidence-based closure and useful delegation up to 20.

The file

Show its rules and text

We split this skill into the rules it gives an agent

15 rules. About 0 are common, 13 are distinctive, and 2 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈13

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 2

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Stop early only when a tool forces it or context is nearly full.

    line 21
  2. Distinctive:

    Always ask the owner first for: protected issues (security, hardware, resilience, P0/P1), accepting a material risk, closing a draft or a …

    line 39
  3. Distinctive:

    Read the repository's AGENTS.md, contribution guide and known-live-issues index.

    line 48
  4. Very specific:

    Create the ledger from the ledger template in an ignored directory, such as out/backlog-burn/ledger.md.

    line 50
  5. Very specific:

    Call orchestrator_capabilities once.

    line 63
  6. Distinctive:

    Do not watch it yet.

    line 101
  7. Distinctive:

    Use up to 20 delegates for the whole run, repairs and reviewers included, under the capacity protocol.

    line 123
  8. Distinctive:

    Start one implement lead per issue in its own worktree.

    line 125
  9. Distinctive:

    Never implement an issue fix in this thread.

    line 132
  10. Distinctive:

    Ask every decision the code cannot settle together through the question tool, per the question playbook: up to 10 per round, one decision …

    line 145
  11. Distinctive:

    Every blocker becomes a question: access, approval paths, outside parties.

    line 151
  12. Distinctive:

    After two failed attempts to diagnose missing access, stop and ask for it.

    line 152
  13. Distinctive:

    When answers arrive: record them, dispatch unblocked work, run approved live actions through the host's approval path and close confirmed …

    line 156
  14. Distinctive:

    Not zero: every remaining item is owner-held, in flight elsewhere or waiting on a submitted question.

    line 189
  15. Distinctive:

    When changing this workflow, exercise the validation cases.

    line 198

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Burn the backlog to zero

Close delivered or obsolete issues and PRs with evidence; run fixes while triaging.

Apply plan-report at proposal, task completion, final handback and on request. Review/fix rounds get short chat updates. The main thread presents the report.

The turn rule

  • Each turn completes steps 2–7 before watching or yielding.
  • watch_pull_request tells you to end your turn. So call it last, in step 8, after the pass. Never watch a PR and then stop with untriaged items, free slots or unasked questions.
  • A watch wakes you only for that PR. Triage the issues this turn.
  • Stop early only when a tool forces it or context is nearly full. Write the ledger's resume record first and name the next step.

Scope and authority

Issues and PRs share one scope: the current repository, unless the owner names an organisation. For an organisation, read org mode and follow it instead of this file.

  • A burn-down request ("clear", "burn down", "get to zero") authorises: implementing through delegates, pushing scoped branches, opening PRs, requesting independent review, squash-merging when the repository's gates are green, babysitting unowned open PRs, closing PRs unmerged under the PR rules, and closing ordinary issues under the classification rubric.
  • Merging includes any automatic post-merge deployment; name it in the report.
  • "Inventory", "triage only" or "ask me" requests propose actions only: no closures, dispatch or merges.
  • Always ask the owner first for: protected issues (security, hardware, resilience, P0/P1), accepting a material risk, closing a draft or a PR whose changes exist nowhere else, and every manual deployment, live host, secret or destructive-data action (one approval per action).
  • Explicit read-only, no-merge, no-close or release holds override all of the above. Never bypass a required approval or force-push a shared branch.

1. Set up (once per run)

  1. Read the repository's AGENTS.md, contribution guide and known-live-issues index. For careynas.net, also read its profile.
  2. Create the ledger from the ledger template in an ignored directory, such as out/backlog-burn/ledger.md. Confirm with git check-ignore.
  3. Inventory with these exact commands (repeat them on every wake):

    gh pr list --repo OWNER/REPO --state open --limit 1000 \
      --json number,title,author,headRefName,headRefOid,isDraft,updatedAt,url
    gh issue list --repo OWNER/REPO --state open --limit 1000 \
      --json number,title,labels,assignees,updatedAt,url

    Record both counts and every ID. If a connector tool fails twice, use gh.

  4. Call orchestrator_capabilities once. Record the model, effort, providers and any concurrency limit in the ledger header.

2. Triage every item from headers

Give every open PR and issue a provisional class from its title, labels, draft flag, linked PRs and dates. Do this for the whole list in one pass and write it to the ledger. Headers only; no bodies yet.

  • PR classes: in-flight-elsewhere, adopt, retire, ask.
  • Issue classes are listed in the ledger.
  • Mark issues covered by an open PR as in-flight and link the PR.

Steps 3 and 4 check each item before you act.

3. Settle each PR's owner

For each open PR, decide whether someone else is actively working on it:

  1. gh pr view N --json commits,comments,reviews,updatedAt,autoMergeRequest.
  2. t3_thread_search with the PR number (for example #4540), then with its head branch. For each hit, t3_thread_read it.
  3. The PR is in flight elsewhere when either is true:
    • a T3 thread that is not archived or settled has a pending watch on this PR, is running, or acted on this PR's current head in the last 24 hours;
    • a person commented or reviewed on the current head in the last 24 hours. Agents often post under the owner's account; a comment or PR body marked as agent-written is not a person.
  4. Otherwise it is unowned. An assignee, an old branch, a recent push, bot activity or enabled auto-merge alone is not ownership.

Record the evidence. For in-flight-elsewhere PRs, record the owner link and add no watch, push or review request. Classify unowned PRs as adopt, retire or ask under the PR rules.

Load babysitter before you act on the first adopted PR. Do the actionable work now: update a stale branch, fix failing checks through a repair delegate, answer review threads, merge when gates are green. Do not watch it yet.

4. Check issues and close what is done

Work through issues in batches of about ten, writing the ledger after each batch. For each issue, read the body and comments, then the minimum current code or docs needed. Apply the classification rubric:

  • Already done, obsolete or not worth the remaining work: close it with an evidence comment and read back the state. Protected issues become questions.
  • Missing evidence: investigate in the code first. Ask only when the evidence is out of reach.
  • Live work: split the source change (may be dev-ready) from the exact live action (becomes an approval question).
  • Dev-ready: add it to the ready queue and go to step 5 as soon as a slot is free. Do not finish triaging every issue before dispatching.

5. Dispatch ready work

Load implement and ship before the first dispatch, and run their preflight once.

  • Use up to 20 delegates for the whole run, repairs and reviewers included, under the capacity protocol.
  • Start one implement lead per issue in its own worktree. The brief names the repository, base, issue, acceptance criteria, authority, child-slot budget and required skills.
  • State the main thread's exact model and effort in each brief for leads and coding workers. Reviewers follow review's selection and eligibility rules. Owner choices override both.
  • Record each returned task ID and worktree; no receipt means not dispatched.
  • Never implement an issue fix in this thread. If delegation is unavailable, ask the owner. If a lead cannot start its own reviewer, get the independent review yourself against the exact committed head; self-review never counts.
  • The main thread owns every PR link, watch and merge. Delegates never watch.
  • One issue, one PR. Use Refs #N instead of a closing keyword when acceptance needs live verification or the issue is protected.

On a delegate result: integrate, open or update the PR under ship, and refill the slot.

6. Ask the owner, in one batch

Ask every decision the code cannot settle together through the question tool, per the question playbook: up to 10 per round, one decision each, recommendation, tradeoff, issue numbers and evidence. Do not ask what the code, issue, comments or earlier owner answers already settle. Dispatch independent ready work before you ask a question that pauses the turn.

Every blocker becomes a question: access, approval paths, outside parties. After two failed attempts to diagnose missing access, stop and ask for it; keep local implementation and review moving meanwhile. Only the owner can declare a hold. Silence is not approval.

When answers arrive: record them, dispatch unblocked work, run approved live actions through the host's approval path and close confirmed issues.

7. Check before ending the turn

Re-run the step 1 inventory commands. Then answer each line in the ledger:

  1. Any PR or issue not yet triaged? Triage it.
  2. Any adopted PR with work you can do now? Do it.
  3. Any ready issue and a free slot? Dispatch it.
  4. Any blocker without a submitted question? Ask it.
  5. Any closure not read back? Read it back.

Continue until all five answers are "no".

8. Watch, then end the turn

Now call watch_pull_request once for each adopted PR that is waiting on checks or reviews. Skip PRs already watched and in-flight-elsewhere PRs. Then end the turn with a status message containing these counts:

PRs: N open, N triaged, N adopted (N watched), N in flight elsewhere, N closed, N merged
Issues: N open, N triaged, N closed, N ready, N dispatched (task IDs), N awaiting answers
Next: <what wakes this run and the first action on wake>

On wake, repeat the inventory and resume from the ledger.

Finish

  • Zero: a fresh inventory shows no open issues or PRs, and every closure has verified delivery or authorised retirement.
  • Not zero: every remaining item is owner-held, in flight elsewhere or waiting on a submitted question. Name each with its owner link and trigger.
  • Not finished: anything else. Keep working.

A closed PR never satisfies its linked issues. At zero or not zero, finish ship's consolidated retrospective, then report per the ledger's handoff, keeping local checks, CI, merge, deployment and live acceptance separate.

When changing this workflow, exercise the validation cases.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.