Home / All Skills / Workflow & Core Tools / continue-claude-code-work
English (EN)简体中文 (ZH)日本語 (JA)
Workflow & Core Tools · daymade/claude-code-skills

continue-claude-code-work

Continues interrupted Claude Code work only after local history has been read and reconciled. Use when the user provides a Claude Code Session ID, asks to pick up prior Claude work, says a Claude run was interrupted or compacted, or wants the current Agent to take over without `claude --resume`. Reconstructs the original business outcome, unfulfilled requests, user corrections, successful prior assets, current workspace truth, and the next action that directly advances the goal. For Codex sessions use continue-codex-work.

L2 Local · Restricted workspaceV1 Commit-PinnedWrite FileDual-pass Audit
Author daymadeGitHub daymade/claude-code-skills ↗Stars 1385License MIT（仓库根 LICENSE：Copyright (c) 2025 daymade；GitHub API spdx MIT）Commit d5c4678cb5
Control levels describe skill capabilities and suggested control measures; they do not represent quality or safety certifications, nor do they dictate usability. Usage decisions rest entirely with users and administrators. Different agent hosts apply varying constraints; do not assume hosts provide automatic protection. Third-party skill copyrights belong to their original authors. Use at your own risk.

1How It Works & Mechanism

It functions as a 'continuation orchestrator' containing zero standalone scripts or file operations: all capability stems from the process discipline in SKILL.md. Step one strictly mandates invoking the sibling skill `daymade-claude-code:read-claude-code-history` to obtain a 'verified read receipt', prohibiting any project modifications prior to receiving it.

daymade-claude-code/continue-claude-code-work/SKILL.md
Invoke `daymade-claude-code:read-claude-code-history` for the exact Session. Ask that Skill for a chronological evidence briefing, not separate “last user” and “last assistant” buckets.
Depends on another skill in the same repository (`read-claude-code-history`, responsible for parsing local session databases); this skill's directory contains only a single SKILL.md file.

The read receipt has a fixed checklist (Session ID/project, source & temporal coverage, compaction boundaries, every retained human request, end reason, pending calls, modified files, and explicit gaps); if gaps might contain the goal, it requires re-reading using the reader's full mode; it halts immediately if identity or provenance is ambiguous.

daymade-claude-code/continue-claude-code-work/SKILL.md
If identity, ancestry, or authorship is unresolved, stop; continuing from a guessed Session is worse than asking for the missing evidence.
Fail-closed posture: prefer asking the user over continuing from a guessed session.

Step two compresses the evidence into a 7-field 'continuation contract' table (original business outcome, current explicit requests, completed, unfulfilled, user corrections, usable assets, and immediate next action), where each field requires independent evidence—prompts like 'continue' do not count as specifications on their own.

daymade-claude-code/continue-claude-code-work/SKILL.md
| Next direct action | One action whose success visibly reduces the unfulfilled business result |
The table mandates distinguishing between 'claimed completion by the previous agent' and 'independently verified completion currently'.

Step three reconciles with current reality: reads the target project's current AGENTS.md/CLAUDE.md alongside authoritative state (historical instructions may be obsolete), audits cwd/branch/working tree, and verifies whether writes, commits, or background tasks recorded in history actually landed—transcripts record attempts, not reality.

daymade-claude-code/continue-claude-code-work/SKILL.md
A transcript records attempts as well as successes.
Prevents treating 'ran last time' as 'still present now'.

The execution gate is the goal-reduction test: before each directional change, ask 'If this succeeds, which original unfulfilled result becomes smaller?' Any branch failing to answer is logged as a proposal and discarded in favor of returning to the mainline; `claude --resume`/`--continue` are explicitly banned.

daymade-claude-code/continue-claude-code-work/SKILL.md
Before changing direction, ask mechanically: “If this succeeds, which original unfulfilled result becomes smaller?”
Strictly restricts 'continuation' to advancing unfinished business outcomes, preventing drift into summarization, polishing history tools, or launching new experiments.

Concludes by reporting against the contract: Recovered goal / Executed / Now complete / Remaining / Not repeated; explicitly forbids claiming 'continued successfully' when only context extraction or workflow steps have been completed.

daymade-claude-code/continue-claude-code-work/SKILL.md
Do not say “continued successfully” when only context extraction or a process step completed; the business result is the completion unit.
The unit of completion is business outcome, not successfully loading a save state.

2Core Capabilities

01Prerequisite read receipt gate: no verified history brief, no action taken
027-field continuation contract reconstruction (business outcome / requests / completion / corrections / assets / next steps)
03Reality reconciliation: verification of AGENTS.md / branch / working tree / external action landing
04Goal-reduction test to prevent drift + strict ban on --resume/--continue
05Contract-based reporting (including 'Not repeated' rejection paths)

3External Dependencies

KindDependency / Endpoint
cliclaude（宿主 CLI，明确禁止 --resume/--continue）
skillread-claude-code-history（同仓兄弟技能，读本地会话历史）

4Recommended Governance & Evidence

Suggested Controls for Safe Operation
For Agent Runtime
Restrict working directory
For Administrators
Retain operation logs
Evidence & Triggers (File Paths & Excerpts)
  • 本地工作目录: File I/O and artifact output: (Normal project operations executed by the agent under authorization—the skill itself contains no write directives; read receipts and contracts are mandatory prior to changes)
Audit Notes: Pure SKILL.md orchestration: no scripts, no network access, no credential reading, and no proprietary write routines. Functionality = first reads local session history via a sibling repository reader skill (read-only for local data), then drives the agent to continue unfinished business tasks in the current working tree (writes occur under normal agent authorization). Note: read content includes the user's full past sessions (privacy-sensitive); evidence reliability hinges on the quality of the dependent reader skill.

5Second-Pass Verification

  • [ok] 无脚本/无网络/无凭证 — 目录仅 SKILL.md；内容不含任何命令执行或网络指令
  • [ok] 引用的兄弟技能存在 — read-claude-code-history / continue-codex-work 均在 daymade-claude-code/ 下有 SKILL.md
  • [ok] read receipt 门禁是真实前置（非可选） — Step 1 标题 Obtain a verified read receipt；Do not act until the receipt states
  • [ok] 功能声明 vs 能力相符（不夸大） — 宣称‘continues work only after history read and reconciled’——SKILL 全文严格围绕该承诺，无越界声称
  • [ok] 本技能不自己解析历史 — 解析/搜索全部委托 read-claude-code-history（含 full mode 重读），与目录内无代码事实一致

6Verdict & Scenarios

  • Standardizes 'continuing interrupted Claude Code work' into a verifiable workflow, eliminating common failure modes (acting on memory, treating summaries as completion)
  • Separates evidence from claims: transcripts are records of attempts; completion status must be independently validated now
  • Goal-reduction test serves as a mechanical gate against scope drift
  • Fail-closed: stops whenever identity or lineage is unclear, preferring to ask the user
    • Best for: Ideal when an interrupted or compacted Claude Code session needs to be transferred to the current agent for continuation (e.g., user provides a Session ID or says 'continue previous work'), requiring strict grounding in actual history without repetition or drift. Also well-suited as a workflow template for other continuation skills.
    Not for: Not suitable for Codex sessions (use continue-codex-work instead); not suitable when native `claude --resume` has already restored context in the current environment (SKILL explicitly states no re-reading needed); not suitable for users with no readable history or who only want a lightweight summary (this is a continuation executor, not a summarizer).
    Install Agent-Ready
    ① Verified Mirror 2026-09-09
    Option A · Direct DownloadDownload continue-claude-code-work.tar.gz
    sha256: 8c7c7550b0a7ecaf…
    Option B · Copy Guide for Agent
    install.json
    Agents automatically download and verify mirrors from this JSON guide.
    ② Upstream GitHub
    GitHub Repo ↗
    Pinned to commit d5c4678cb5.