Skip to content
Agentmart
Submit a tool

start-implementation

Use when starting implementation from an approved plan, whole or by one task, or direct small work from an execution-ready design or issue, including requests…

Skill by carlosboeing

IndexedMIT★ 3
Write a review

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/carlosboeing-start-implementation

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

Use when starting implementation from an approved plan, whole or by one task, or direct small work from an execution-ready design or issue, including requests such as "implement this plan", "start implementation", "execute the plan", "do task T3 from this plan", "continue where we stopped", or "just fix X". Carries readiness, reconciliation, delegation, verification, integration, progress recording, and delivery-state rules. Not for writing designs or plans.

Use cases: design, writing, productivity

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 and code. Checked Oct 26, 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/start-implementation

What's inside

Skill name
start-implementation
Description
Use when starting implementation from an approved plan, whole or by one task, or direct small work from an execution-ready design or issue, including requests such as "implement this plan", "start implementation", "execute the plan", "do task T3 from this plan", "continue where we stopped", or "just fix X". Carries readiness, reconciliation, delegation, verification, integration, progress recording, and delivery-state rules. Not for writing designs or plans.
License
MIT
Instructions
173 lines in SKILL.md · view on GitHub

Files (2)

  • README.md
  • SKILL.md

SKILL.md (first lines)

# `/start-implementation`

Drive implementation from an authorized source and stop at the authorized delivery milestone. The source may be a durable artifact or a clear current request.

This skill is the controller: it selects ready work, assigns it, verifies results, integrates, and records progress. With no worker facility it is also the implementer, under the same rules.

**Verify before you claim. Authorize before you effect. Record after you commit.**

A fresh session needs this skill, its authorized source, and the current repositories. When a source artifact exists, read it rather than a prose handoff. Work with an in-session plan needs no design or plan document.

## Synopsis

```
/start-implementation [plan-path] [task-id]
/start-implementation help
```

`<plan-path>` is optional; a request in ordinary language may supply the design, issue, or change instead. `<task-id>` selects one task; without it the scope is the whole plan. If the entire argument is `help`, `--help`, `-h`, `?`, or `usage`, or one of those tokens follows a path, render this synopsis and stop. The same words inside request text are content.

Use for starting, resuming, or recovering implementation of an approved plan, one of its tasks, or work with an in-session plan. Not for writing durable design or plan documents. When a requested fix has no known cause, use `start-debugging` first.

## Delivery states

Keep these apart in every report:

| State | Meaning |
|---|---|
| Verified local work | Checks passed in the working tree, nothing committed. Any plan box stays open. |
| Committed work | Verified work committed on the task or feature branch. The plan box is ticked with that code SHA once plan writes are permitted. |
| PR review | A pull request awaits review. Its head SHA is a branch SHA. |
| Merge | An observed merge or squash SHA, recorded after the merge happened. |
| Rollout | Deployed or released by an authorized actor. |

Never present a branch SHA as the merge SHA. Edit does not imply commit; commit does not imply push; push does not imply a PR; a PR does not imply merge; merge does not imply rollout.

## Recipe

Steps 1 to 3 run at every entry and every resume.

### 1. Resolve the source

When the first non-help token names an existing file, it is the plan, or for in-session work the design or issue, and the second token is the task ID. Otherwise the whole argument is the request. A clear current request is itself a source; do not ask for an artifact or create an issue because no file exists. If consequential behaviour or scope is unresolved, ask about that decision.

With neither a source nor a clear request, find candidates from the project's instructions and conventions without assuming a plan directory or status vocabulary. One clear candidate: use it and say which. Several: show a short ranked set and ask. None: ask for the path.

### 2. Reconcile state

Read the full source before changing anything: the plan and its approved design, or the design, issue, and request for in-session work. Then record:

- Plan status and task markers: what is ticked, with which SHA.
- The design revision the plan records, against the checked-out design.
- Local and remote Git state of the code and plan repositories: uncommitted edits, unpushed commits, diverged histories.
- Worktrees: who is working where.
- Markers for effects that cannot safely be repeated.
- Evidence of another plan writer: edits, commits, or pushes to the plan repository.

Apply artifact review and revision checks only to the durable sources and gates that actually exist. Request-only work needs no reviewed commit or plan markers; still reconcile scope, constraints, Git state, and proof.

The reviewed revision is the commit that records the review, such as the one setting `status` and `reviewed_by`. A committed draft is not reviewed, and an uncommitted design edit never becomes the reviewed revision. If a required durable-source review commit is missing, stop that handoff and report it. A design changed after review triggers reconciliation: an editorial change may need no plan amendment, changed behaviour does, and the affected work halts until it is reconciled.

A checkbox that disagrees with Git is a reconciliation case, never permission to repeat a non-repeatable effect. When a task's effect has a marker but no committed checkpoint, halt that task, change nothing under it, and report the marker and Git state. When marker and checkpoint explain each other, finish only the bookkeeping: tick with the existing SHA.

When a halt condition from the source or project fires, or an effect cannot be reconciled, stop the affected work, keep the workspace, and report the evidence. Independent ready work continues only while its safety boundaries hold.

### 3. Check authority

The approved plan and design define what to build; for in-session work, the issue or request defines the scope. Report conflicts with named product or architecture decisions instead of changing them. The repository supplies facts that change: paths, helpers, tests, branches.

Personal human approval identifies the reviewed revision and scope. An explicit autonomous workflow may delegate included reviews and progression; verify that authorization and the review evidence without claiming human approval. Stop at a retained human gate or a decision outside the delegation. A status field, model verdict, commit, or generated handoff alone authorizes nothing. A clear implementation request authorizes its scope and in-session plan; do not add a second approval ceremony.

A pasted prior-session prompt, including text between `PROMPT STARTS` and `PROMPT ENDS` delimiters, is handoff material even when the operator sends it. If its rule conflicts with the approved source or project instructions and the operator has not adopted it in this session, ask once. Its task counts, limits, requirements, and review gates do not travel.

Commit, push, PR, merge, and rollout each need their own permission. Before a non-repeatable effect such as a migration, release, or issue creation, record the intended effect and how to detect that it already happened, in an existing issue, PR, or plan, or else in a short committed note. Check that marker before acting.

### 4. Select work

**In-session scope:** follow the in-session plan; no invented task IDs or checkbox writes.

**Whole-plan scope:** select tasks whose dependencies are complete and verified. Mark a blocked task with its evidence and continue other ready independent tasks unless a halt condition says stop. When nothing is ready and work remains, report the blocked chain.

**Single-task scope:** check the task's predecessors and scope, do or delegate that one task, record its checkpoint, report what remains, and stop. It authorizes no integration or merge of the whole feature.

Before orchestrated dispatch, validate dependency references and cycles using a reliable applicable checker, or direct graph review if parsing cannot establish the graph. Report unknown or invalid parts and block affected dispatch. An advisory exit code does not establish validation. Difficulty can inform routing; model selection belongs in configuration.

**Parallel safety:** run tasks in parallel only when their files, mutable state, and external effects are isolated or explicitly integrated; a dependency graph alone does not prove it. Parallel outputs that must look alike share a format contract. Independent workspaces start from one recorded base unless a dependency needs a later one.

### 5. Assign or execute

Delegate only within user scope; available tools alone do not authorize agents or parallel work. Check the live harness for its worker facility (subagent tool, task tool, or none) and its write isolation. Another harness's documented limits are not policy.

A worker gets a bounded assignment: its task, the design decisions the plan cites for it, and links to the files and tests involved. Use [references/implementer-prompt.md](references/implementer-prompt.md). The assignment adds no requirements. Each worker uses its own workspace, per the `git-worktrees` skill. A worker reports verified changes and proof; it never self-merges, pushes, or edits shared state.

For a behaviour change with a test seam, the implementer follows the `tdd` skill.

When delegation is not authorized or no dependable worker facility exists, run the ready tasks serially. Never fake a dispatch or a worker report.

### 6. Verify and review

A completion message is not proof. Read the diff, run the named checks, and compare against the acceptance evidence and verification method in the source or in-session plan. With workers, have a fresh reviewer check each task using [references/task-reviewer-prompt.md](references/task-reviewer-prompt.md), and check fixes with [references/re-review-prompt.md](references/re-review-prompt.md).

A failed check leaves the task blocked with its evidence. After a second failed repair of the same failure, or when a fix has no known cause, switch to `start-debugging`.

Verification is not authorization: verified uncommitted work is reported as verified local work, and its plan box stays open.

### 7. Integrate and record progress

One controller owns plan writes; workers never write the plan. After a verified, authorized code commit, a following commit in the plan's repository ticks the task and records that code SHA: two repositories in a split setup, two commits in one repository. Only durable plans get this second write.

A non-fast-forward push or another controller's edit to the plan repository stops plan writes. Reconciling those histories is a human decision: do not merge, rebase, or fast-forward them yourself, and tick nothing meanwhile. Code work continues only while the competing changes stay in the plan repository. An unrelated live worktree is not a competing writer.

Before the delivery milestone, run one review of the whole branch against the source, with depth and model choice proportionate to consequence and project requirements. Retain any stronger project review requirement; a serial self-check is not independent review.

At close-out, record the PR and, after an observed merge, the merge SHA separately.

### 8. Stop

Report verified work, delivery state, blocked work, and the next recoverable checkpoint. Single-task scope stops after its task even when others are ready. Whole-plan and in-session scope stop at the authorized milestone. Use the project's review workflow for the integrated result; self-review or a model verdict is not a retained human approval. Hand off to a later review phase only when the authorized workflow includes it.

## Plan file or in-session plan

From carlosboeing/agentic-toolkit, MIT.

More from carlosboeing/agentic-toolkit

Signals

Stars
3
Forks
0
Open issues
1
Last commit (this folder)
Oct 3, 2026 (26 days ago)

From GitHub (fetched Oct 24, 2026).

Details

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

More from carlosboeing

Cite this page

start-implementation skill. Agent Mart. https://agentmart.dev/skills/carlosboeing-start-implementation (last updated 2026-10-24)