Delivery methodology
Comment.io engineering-workflow skills for any AI coding agent (Claude Code, Codex, Cursor) — plan, prototype, build, review, ship. npx skills add comment-hq/skills
npx -y skills add comment-hq/skills --skill delivery-methodologyAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
What its author says it does
Copied from the file, not written here
Choose and enforce a pragmatic engineering delivery topology: a direct PR for independently shippable work or a controlled lift branch for foundational work with unsafe intermediate states. Defines bounded SHA-delta reviews, risk-scaled reviewer counts, review receipts, lift freeze/promotion, and a technical-ready stopping point. Delivery skills call it before planning or shipping; invoke directly when deciding how a large change should reach main.
SKILL.md
7.9 KB, as published. Nobody here has run it
delivery-methodology — ship bounded work without preserving two systems
Read the repo's AGENTS.md (else CLAUDE.md) and any linked delivery/testing
guide first. Repo rules may be stricter. This skill supplies the generic method
when a repo has no equivalent.
Startup operating principle
Optimize for the smallest simple change that gives real users value and yields useful product evidence. Treat complexity as a cost: abstraction, compatibility machinery, enterprise hardening, and exhaustive hypothetical edge-case support need a present requirement or credible failure path. Preserve hard correctness, security/privacy, data-loss, migration, and protocol invariants. Otherwise ship a reversible useful boundary, validate the real user journey, and expand from evidence rather than imagination.
Choose the topology
Use a direct task branch for a complete change that can safely ship in one bounded PR. It starts from the target base and the PR targets that base.
Use a controlled lift branch for foundational work whose partial state is unshippable or whose clean implementation replaces a system across coordinated slices. Record the choice and why before implementation.
Feature flags are a rollout tool, not the default slicing strategy. Use one only around a clean, bounded activation seam. Prefer a lift when a flag would require broad branching, dual writes, duplicated state, compatibility shims, or two authoritative systems.
Controlled lift
- One Project Root/worklog and one
lift/<initiative>own the work. - Slice branches start from the current lift tip and target the lift, never another slice branch.
- Merge slices sequentially with real merge commits. Never squash, rebase, or force-push the lift.
- Merge the target base into the lift periodically and review conflict delta.
- Keep unrelated work out.
- A human-approved lift plan may authorize internal slice merges; promotion to the release branch still follows repo/user merge authority.
- Freeze the lift before promotion. No dependent slice remains open or lands after the promotion starts.
Review explicit deltas
Every target names base_sha, head_sha, a locked subject/scope, acceptance criteria,
invariants, risk tier, and checks. Review base_sha..head_sha plus necessary
impact context—not the accumulated branch by habit. Surrounding code is evidence,
not a broader audit surface. A finding is in scope only when the target delta
introduced or changed it, or when the delta makes an existing path violate an
explicit acceptance criterion or invariant.
- Routine: one reviewer.
- Sensitive: two independent lenses for authorization, migrations/storage, protocol/compatibility, concurrency, native code, destructive behavior, or credible data-loss risk.
- Exceptional: three only for exceptional blast radius.
Gather one complete finding batch, fix compatible findings together, and run focused validation. Review later fixes from the last reviewed head plus any invalidated invariant. Normal work gets at most two finding-bearing rounds; then redesign, prove the disputed case, record residual risk, or ask a human.
A blocking finding needs a concrete correctness, security/privacy, data-loss, compatibility, migration, protocol, or stated-requirement failure. Decline style preferences, speculative refactors, unrelated debt, and unsupported hypotheticals with concise reasoning.
For a genuinely unrelated defect, do not expand the current review or branch.
Search the GitHub issue ledger and attach evidence or create one focused issue
(file-bug when available), link it in the receipt, and continue. If it is
actively release-breaking, creates a credible security/privacy or data-loss
emergency, or makes the current delivery unsafe to evaluate, stop and ask the
owner human to start a separate worktree job. A defect exposed or caused by the
target delta remains in scope.
Emit a receipt
Record:
kind: slice | main-sync | direct | promotion-fix
base_sha: <commit>
head_sha: <commit>
merge_sha: <commit that entered the lift, or head_sha for direct>
source_sha: <second parent for main-sync; omit otherwise>
scope: <bounded change>
invariants: <properties checked>
risk: routine | sensitive | exceptional
reviewers: <count/lenses>
checks: <commands/evidence>
findings: <fixed or declined with reasons>
residual_risks: <explicit remainder>
For a slice merge require base_sha == merge_sha^1 and
head_sha == merge_sha^2. For a main sync require
base_sha == merge_sha^1, source_sha == merge_sha^2, and
head_sha == merge_sha. A direct or promotion-fix receipt must bind a
single-parent commit with base_sha == merge_sha^1 and head_sha == merge_sha.
Ordered receipt merge_sha values must exactly cover the lift's first-parent
chain from its recorded base to frozen head.
The bundled scripts/verify-lift-receipts.sh checks a TSV ledger:
kind<TAB>base_sha<TAB>head_sha<TAB>merge_sha<TAB>source_sha-or--
Run:
<skill-dir>/scripts/verify-lift-receipts.sh <lift-base-sha> <frozen-head-sha> <receipts.tsv>
Test the right boundary
- During implementation/review fixes: narrowest useful checks.
- Lift slice: focused checks for its explicit delta; record them in the receipt.
- Direct release candidate: complete affected lane once on the frozen commit.
- Lift promotion: complete affected lane once on the frozen lift head plus the initiative's realistic integration/migration scenario.
- Release merge queue: full matrix on the speculative merge.
Do not claim a test command is slice-scoped when its selector compares with the release branch. Run independent final gates concurrently where practical.
Finish pragmatically
A boundary is complete when acceptance passes, required evidence passes, no known severity-blocking defect remains, mandatory repo rules hold, and residual risks/follow-ups are recorded. Optional polish, speculative enterprise readiness, exhaustive hypothetical edge cases, and unrelated cleanup become follow-ups. Phases organize implementation; they are not automatic review, staging, steering, or artifact gates.
Update the worklog only on material state changes. Create separate Plan/design/ ADR docs only when they provide a real human review surface or durable decision.
Distinguish implementation-converged, candidate-certified, PR-technically-ready, human-authorized, and merged. Do not keep an active session alive indefinitely waiting for an external reaction the repository does not require.
Promote a lift
- Merge every intended slice and freeze the lift.
- Merge current target base into the lift with a real merge commit.
- Review and receipt conflict resolutions/uncovered delta.
- Mechanically verify ordered receipt coverage to the frozen head.
- Run final affected and integration certification on that SHA.
- Run one composition review: receipt coverage, cross-slice seams, migration/ cut-over order, single-source-of-truth outcome, and residual delta.
- Open one promotion PR from the certified SHA. A draft is allowed before review only when the sole official reviewer requires a PR; keep it draft until that same SHA passes.
- Merge through the target branch's normal mechanism when explicitly authorized, then verify every slice head and frozen lift head reached it.
Domain workflows such as production deployment, incidents, storage migration, or protocol cut-over may impose stricter irreversible-risk gates. Preserve them without reintroducing generic repeated review ceremony.