AGENT SKILL

deliver-feature

Use when a feature should move through the whole pipeline. The user asks to "deliver this feature", "build this feature end to end", "continue the feature", or "run the pipeline". Also when a session resumes work that has existing spec or slices artifacts, or nobody remembers where a feature stands.

License 许可
MIT
所属仓库
adriancodes/skills
最近更新
2026-11-12
Stars
★ 1
Forks
0

01 前言 Front Matter

name
deliver-feature
license
MIT
metadata
{ "summary": "Runs one stage of the feature pipeline per invocation by reading its artifacts, selecting the right skill, and stopping at the stage gate.", "category": "Building" }
description
Use when a feature should move through the whole pipeline. The user asks to "deliver this feature", "build this feature end to end", "continue the feature", or "run the pipeline". Also when a session resumes work that has existing spec or slices artifacts, or nobody remembers where a feature stands.

02 资源文件 2

03 SKILL.md 正文

Deliver Feature

Overview

Move a feature through spec, task creation, implementation, and verification, one stage per invocation. Detect the stage from files, not conversation memory; run the matching skill; stop at its gate. It routes work; it does not replace confirmed plans.

When to Use

  • "deliver this feature", "ship this feature", "continue the notifications work", "what's next on exports", "run the pipeline"
  • A fresh session resumes a feature whose spec or slices files exist
  • The feature's state is unclear: the artifacts, not recollection, should say

Do Not Use When

  • The user wants one specific stage: invoke that skill directly (create-spec, create-tasks, implement-task, verify-work)
  • A quick fix with no pipeline artifacts and no need for them: just do it
  • Running unattended on a schedule: design that wrapping loop with build-loop; the conductor stays one-invocation-one-stage inside it

Required Context

  • The feature's topic, enough to find its artifacts
  • The topic's decision log and *-slices.md, read in full.
  • Repository instructions that re-home docs/specs/.
  • Frontmatter status and every slice tick.

Workflow

  1. Read the baton. Find and fully read the decision log and slices file. Include frontmatter. Never infer state from filenames. If two open specs match, ask which. Treat a missing file as state and quote it as missing.

    Done when each artifact is classified as missing, open, confirmed, complete, or invalid.

  2. Detect the stage: artifacts only. First matching row, top to bottom:

    Artifact state Stage to run Exit at
    Malformed artifact, unknown status:, slices without a linked spec, or a confirmed zero-slice file Invalid artifact: stop and name the exact contract violation Artifact corrected and reconfirmed in a later invocation
    No decision log, or status: open Spec: create-spec: capture mode only for answers gathered live in this conversation; an open spec's own contents never count as "the answers" The spec's confirmation gate
    Spec confirmed (or confirmed-by-override), and no slices file or slices status: open Task creation: create-tasks The tasks read-back (slices status: confirmed)
    Slices confirmed with an unticked slice whose blockers (if any) are all ticked Implementation: implement-task on the first such slice That slice ticked
    Slices confirmed, unticked work remains, and no slice is unblocked Invalid graph: stop and report the cycle, missing blocker, or unsatisfied edge Slices file corrected and reconfirmed in a later invocation
    All slices ticked, ## Verification section empty Verification: run verify-work read-only on the feature's artifacts If clean, write the report summary + date into ## Verification; if findings remain, leave it empty and request explicit fix authority
    ## Verification filled Done: summarize the paper trail (spec, slices, verification) and stop :

    Treat status: open as a stop signal. Run that artifact's stage, however finished it looks. Treat confirmed-by-override as confirmed.

    Done when the stage is named before work starts.

  3. Run the owning skill. Invoke the sibling and follow it fully. Never inline a shortcut when the skill is installed.

    When it is missing, name the gap and installation path. Stop unless the user approves the fallback:

    • spec → ask one decision question at a time and confirm a written log;
    • tasks → write vertical, demoable, session-sized slices with blockers;
    • implementation → build one slice test-first, run its demo, and tick it;
    • verification → execute hostile cases against the artifact.

    Done when the stage's own gate is reached.

  4. Exit clean. Report the stage, outcome, and current artifact state. Name what the next invocation will do. Then stop. Run exactly one stage per invocation, including planning stages.

    Done when the report ends with the next-stage line.

Example: a resumed session

Read: spec 2026-07-05-notifications.md: status: confirmed, 6 decisions. Slices file: 1 ✓, 2–4 open, all blocked only by 1. Stage: implementation: task 2 is the first unblocked. Running implement-task on it. Exit: slice 2 ticked. Next invocation: slice 3.

Common Rationalizations

Excuse Reality
"The slices are small and related: faster to do all three in one pass" Faster until the third slice inherits two slices of stale context. One stage, one invocation; fresh windows are the pipeline's whole trick.
"I can see what's needed from the file names: no need to re-read the log" Skimmed specs are how confirmed decisions get rebuilt differently. Both files, in full, first.
"Slice 1 established the patterns; the rest is mechanical: low risk to batch" "Mechanical" is a claim the demo criterion tests one slice at a time.
"'Continue shipping' means they want a finished feature, not a status update" It means advance the pipeline. The stage report is progress, resumable by anyone tomorrow.
"Actually, websockets would be cleaner than polling here" status: confirmed closes that. Reopening goes through the log's supersede rule with the user (a new numbered entry, their say-so quoted, history intact): never through fresher opinions mid-stage.

Success Criteria

  • Each artifact read in full or its absence established, both states quoted before any work
  • The stage named before it runs; exactly one stage (one slice, if implementing) per invocation
  • Zero confirmed decisions relitigated; zero building on a status: open spec
  • The exit report says what the next invocation will do

Common Mistakes

Mistake Fix
Treating checkboxes as the whole state The spec's status: gates everything; read it first
Re-planning over existing artifacts The conductor routes; changes go through the owning file's rules
Quietly inlining a missing sibling's job Loud degradation: name it, offer install or the one-line fallback, get a say-so
Momentum past an implementation gate One slice per invocation; the user decides to continue

Failure Modes

  • Artifacts contradict: stop at the spec because the upstream file wins. Read back the mismatch before work.
  • No artifacts exist: start at the spec stage. Say so before running it.
  • The topic matches nothing findable: ask for the spec path rather than guessing across features.

Summary

Read the feature artifacts, run the current stage, stop at its gate, and state what the next invocation should do.

04 同仓库更多 Skills

adriancodes
adriancodes/skills

write-agent-instructions

Make the smallest evidence-backed change to persistent project instructions and test downstream behavior.

★ 1
adriancodes
adriancodes/skills

verify-work

Use when a finished, runnable artifact (a script, config, parser, prompt, or CLI, not a diff under review) must survive hostile input before shipping: the user asks to "verify this script", "find the loopholes", "test this properly against bad input", or asks "is this ready to ship" about one. Also when a shipped artifact failed on an input nobody tried, or when verification so far is a happy-path smoke test.

★ 1
adriancodes
adriancodes/skills

understand-codebase

Use when the user asks to "understand this codebase", "how does this feature work in this repo?", "trace this feature through the code", or needs onboarding, an architecture explanation, or answers about unfamiliar code. Also use for read-only code-path questions before a change.

★ 1
adriancodes
adriancodes/skills

tdd

Use when writing or changing production behavior: implementing a feature, fixing a bug whose cause is already known, or altering stored, transformed, migrated, serialized, or externally transmitted data. Also use when a user says "use TDD", "test first", or "red-green-refactor", and when code is being written before a failing test. For a failure whose cause is unknown, diagnose first.

★ 1
adriancodes
adriancodes/skills

simplify-code

Use when the user asks to "simplify this code", "remove the fluff", or "clean up this over-engineered module", or wants working code that is bloated, duplicated, over-abstracted, or dependency-heavy reduced without changing behavior.

★ 1
adriancodes
adriancodes/skills

review-session

Review an agent session for evidence-backed friction and the smallest measurable remedies.

★ 1