Skip to content
Agentmart
Submit a tool

implement

Apply the approved plan on its branch, verify the change against the Definition of Done in both lanes, and mark the pull request ready for review.

Skill by juangcarmona

IndexedMIT
Write a review

Chalk Labs

Need help deploying implement?

Chalk Labs sets up agents like this inside real teams: connected to your tools, tested on your data, with a human in the loop where it matters.

Book a call

Install

Works with Claude Code, Claude apps, Codex, Cursor, Gemini CLI, GitHub Copilot, VS Code, OpenCode, Goose, Windsurf. Read the repository before you give any tool access to your data.

npx agentmarket install skill/juangcarmona-implement

Asks which of your tools to add it to (Claude Code, Codex, Cursor, VS Code, Windsurf, Gemini CLI, Claude Desktop) and shows the change first. See https://agentmart.dev/cli.

Get the weekly diff

The weekly diff: one email a week with new agent tools, notable updates and grade changes, plus security alerts for tools you follow. Double opt-in, one-click unsubscribe.

We use your email only for this newsletter. Privacy

About

Apply the approved plan on its branch, verify the change against the Definition of Done in both lanes, and mark the pull request ready for review. Use when implementing or finishing a planned work item.

No reviews yet. Installed it? Reviews come from verified installers only. How reviews work

Permissions

What this skill can do and what it asks for, from its SKILL.md. Checked Oct 25, 2026.

Tools
No tools found
Secrets
None asked for
Install scripts
None found

Network hosts it talks to 1

  • codeload.github.com · found in install command

Files it touches

  • ~/.claude/skills/implement

What's inside

Skill name
implement
Description
Apply the approved plan on its branch, verify the change against the Definition of Done in both lanes, and mark the pull request ready for review. Use when implementing or finishing a planned work item.
License
MIT
Metadata
author: juangcarmona · plugin: agentic-sdlc
Instructions
67 lines in SKILL.md · view on GitHub

Files (1)

  • SKILL.md

SKILL.md (first lines)

## Purpose

Lifecycle stage 3. Turn an approved plan into a **change that is finished and proven**, and end at the Done gate.

Input is a work item identifier, and it is required. If none was given, ask for one before starting.

Invoke deliberately. No lifecycle stage runs this one on its own.

## When to use

- Implementing or building out a work item whose plan has been approved.
- Finishing a change that is partly done and taking it to the Done gate.
- Re-running the gate after fixing what it reported unmet.

## When not to use

- Designing the solution or deciding the approach. That is `propose`, and its output is the input here.
- Merging the finished change. That is `integrate`.
- Running the deterministic checks alone, outside a lifecycle cycle. That is `verify-like-ci`.

## Workflow

Every external write is shown first and confirmed, except when invoked by an enabled `/ff` run whose durable record authorizes routine lifecycle writes. That mode changes no deterministic check or repository protection rule.

1. **Load the configuration.** Read `docs/engineering-lifecycle.md`. Absent means stop and say to run `adopt` first.
2. **Check the plan was approved.** `read-work-item` plus the pull request's review state via `inspect-pull-request`. In enabled `/ff` mode, the independent review decision recorded for the current plan may satisfy the lifecycle's plan gate. Any approval required by the forge remains required. No applicable approval means stop and point at `propose` for this item. Implementing an unreviewed plan is how a stage builds the wrong thing correctly.
3. **Continue on the plan's branch** via `checkout-branch`. Never a new branch, never a second pull request. The branch already carries the plan; the code joins it.
4. **Apply the plan, task by task**, via `apply-change`, committing incrementally so the branch keeps a narrative a reviewer can follow. Where the configuration's **Provided roles** names a skill for this contract, that skill is `apply-change`; the contract is the same either way.
5. **Surface deviations immediately.** A plan step that cannot be followed as written is reported when it is found, not absorbed silently and mentioned at the end. Update the artifact so the plan and the code stay in step.
6. **Update the documentation the change affects**, in the same branch. Behaviour that changed and documentation that did not is a defect with a delay on it.
7. **Keep the task list honest as you go.** Tick a task only when the code supports it. The authoritative reconciliation against the diff happens in step 9, and it is not yours.
8. **Run the deterministic lane in this session.** `verify-like-ci` runs what the pipeline runs, derived from the pipeline definitions rather than memory. `verify-runtime` captures evidence for a user-facing change. Their output belongs here, where a failure is immediately actionable.
9. **Delegate the independent audit** to the `done-gate-auditor` agent. It reconciles the plan and its task list against the actual diff, checks that the evidence the Definition of Done requires exists and is linked, names verification that is missing rather than merely failing, and prepares the judgement-lane items with their owning roles. This delegation is required: a session that just wrote the change cannot audit whether the change matches the plan, because it carries its own rationalisations. In enabled `/ff` mode, invoke it with the configured different review model; it may decide only judgement items explicitly marked substitutable in `## Fast Forward`.
10. **Close what the gate reported unmet.** Each item comes with what is missing and who or what closes it. Fix, then re-run the **whole** check, never the sub-command that failed.
11. **Commit and push.** Before staging, confirm the branch is the item's branch. After pushing, confirm the remote matches local.
12. **Mark the pull request ready for review** via `update-pull-request`. This starts CI and notifies people, which is why it is a deliberate act and not a side effect.
13. **Move the state** to what the mapping names for review, via `transition-work-item`.
14. **Harden the process if this stage went red.** A failed pipeline, a check that should have caught something earlier, a step reported done that was not: each becomes one durable guard via `harden-process`. The loop improves while it runs, not at some distant retrospective.
15. **Stop at the Done gate.** **Both lanes, or it has not passed**: a green deterministic lane is half the gate, not the gate. In normal mode, the named role decides judgement items. In enabled `/ff` mode, only the configured different review model may decide items explicitly marked substitutable; every other judgement item remains unmet until its named human role decides it. After the gate, a human reviews the change unless `/ff` has a valid independent-review decision and forge policy permits it. Name `integrate` as what follows once the applicable review passes, and do not invoke it.

## Validation

- The plan was approved before any code was written.
- Work happened on the plan's existing branch, with no second branch or pull request.
- Every deviation from the plan is recorded in the artifact, not just in the conversation.
- `done-gate-auditor` ran, and its reconciliation of plan against diff is what the gate rests on.
- Every applicable Definition of Done item has a verdict; unmet ones are named individually with their owning role.
- No check reported as passing was actually unrunnable.
- Documentation affected by the change was updated in the same branch.
- The pull request is ready for review and the remote matches local.
- Anything that went red produced a durable guard.

## Common pitfalls

| Pitfall | Fix |
|---|---|
| Implementing a plan nobody approved | Stop and point at `propose` |
| Opening a second branch or pull request | One branch, one pull request, one item |
| Auditing your own work against the plan | Delegate to `done-gate-auditor`. Independence is the point |
| Reporting a subset of checks as green for the whole | A chained command stops at the first failure: the later checks never ran, and their silence is not a pass |
| Treating `unrunnable` as `passing` | A check this machine cannot run leaves its Definition of Done item unmet, by name |
| Re-running only the check that failed | Re-run the whole Definition of Done |
| Softening the gate to unblock the change | Widening it is not passing it |
| Deciding a judgement-lane item yourself | The named role decides. Never approve your own work |
| Running `integrate` after the gate passes | The stage ends at its gate |

From juangcarmona/agent-toolkit, MIT.

More from juangcarmona/agent-toolkit

Signals

Stars
0
Forks
0
Open issues
0
Last commit (this folder)
Oct 19, 2026 (11 days ago)

From GitHub (fetched Oct 24, 2026).

Details

License
MIT
First seen
Oct 24, 2026
Last checked
Oct 24, 2026

More from juangcarmona

Done for you

Rather not set it up yourself?

Chalk Labs scopes, builds and runs agents for your team: the task audit, the eval harness, the integrations and the weekly ops report. Fixed scope, priced before we start.

See how we build AI agents →Read the case studies

Chalk Labs is a two-founder AI and growth studio. We build AI agents, CRMs and MVPs for Web3 and AI companies, and have shipped 50+ product and token launches.

launches shipped
50+
launches shipped
media placements
500+
media placements
founders on every account
2
founders on every account

Press we've placed: Forbes · CoinDesk · Decrypt · Cointelegraph · The Block

Rahil Jain, Co-founder, Chalk Labs. Ex-Goldman Sachs engineer. You talk to the person who builds it.

Get implement running in your business

Tell us what you want to automate. A founder replies within one working day with a plan, not a pitch deck.

Goes to the Chalk Labs founders, not the newsletter. Privacy

Need help deploying implement?

CallGet help
Cite this page

implement skill. Agent Mart. https://agentmart.dev/skills/juangcarmona-implement (last updated 2026-10-24)