Context Rescue

Create a thread-based recovery summary when an agent loses the thread. Read the relevant conversation and canonical files, summarize what was done, where it...

Security Scan

The skill's instructions, file access, and lack of installs/credentials align with its stated purpose of producing a reorientation summary and do not request disproportionate permissions.

Purpose & Capability

The name/description match the actual behavior: read thread and canonical project files, summarize progress, and write a snapshot. It does not request unrelated credentials, binaries, or external services.

Instruction Scope

SKILL.md narrowly instructs the agent to read the conversation and specific project files (state/*.md, task_plan.md, notes.md), produce a focused summary, and write state/ORIENT.md. It explicitly forbids continuing into implementation, and does not instruct broad data collection or external transmission.

Install Mechanism

This is an instruction-only skill with no install spec or code files, so nothing is written to disk beyond the artifact the skill itself directs (state/ORIENT.md). Low installation risk.

Credentials

The skill requires no environment variables, credentials, or config paths beyond reading/writing project/state files named in the workflow. The requested access is proportional to summarization and reorientation tasks.

Persistence & Privilege

The skill is not always-enabled and does not request elevated or cross-skill configuration. It writes a single snapshot file within the project state, which is consistent with its purpose.

Assessment

This skill is internally consistent and low-risk: it only reads conversation history and specific project state files and writes a short snapshot to state/ORIENT.md. Before installing, confirm (1) the agent will only have access to the threads and project files you want it to read (sensitive messages will be included if accessible), (2) writing state/ORIENT.md won't overwrite important content (back up if necessary), and (3) you are comfortable with the agent producing summaries of any sensitive context it can access. Because it's instruction-only, test it in a sandboxed project or repository first to verify behavior.

Use this skill when an agent is no longer fully sure where the work stands.

This is not for execution. This is for reorientation.

Trigger condition

Use this skill when any of these are true:

• the task feels blurry or fragmented
• too many side-tracks were followed
• the thread is long and the real next step is unclear
• a handoff or resume is happening after a gap
• you need to identify the real next step before doing more work

Core recovery prompt

Read back the full thread content and summarize what we have done so far, where the task is documented, and what the next step is. Do not go further, only produce this detailed summary.

Recovery workflow

1. Read the relevant thread or message history.
2. Read the canonical files:
   • state/HOLD.md
   • state/ACTIVE.md
   • state/DECISIONS.md
   • state/CLOSED.md if present
   • the canonical task_plan.md
   • the canonical notes.md
3. Produce a focused summary with exactly these parts:
   • what we did so far
   • where the task is documented
   • what the actual next step is
4. Write a compressed snapshot to state/ORIENT.md.
5. Stop. Do not execute the next step unless separately instructed.

Output format

## Reorientation summary

### What we did so far
- ...
- ...

### Where it is documented
- `...`
- `...`

### Actual next step
- ...

### Not the next step
- ...
- ...

### Closure criteria
- ...
- ...

state/ORIENT.md format

## Reorientation snapshot
**Thread / Context:** [which thread or task this is]
**What we did so far:**
- ...
- ...

**Where it is documented:**
- `state/ACTIVE.md`
- `projects/<slug>/task_plan.md`
- `projects/<slug>/notes.md`

**Actual next step:**
- ...

**Do not do instead:**
- ...
- ...

**Closure criteria:**
- ...
- ...

**Last updated:** YYYY-MM-DD HH:MM

Rules

• Do not continue into implementation.
• Do not create a new project if one already exists.
• Do not confuse logs with task progress.
• Prefer file truth over vague recollection.
• If the thread and the files conflict, explicitly call that out.

Why this exists

Long threads can consume context and blur the real task. A short, canonical reorientation summary restores focus without pretending the work is done.