How to use the Custom Protocol Replay agent skill (setup guide)

What the skill makes the agent do

An agent skill is a set of instructions and reference files that an AI agent loads when a request matches the skill's description. This one is a downstream specialisation: it applies a fixed method to protocol-reversing work once a parent skill has set the sandbox rules.

The method follows a fixed order. The agent first maps the session's state machine (handshake, negotiation, authentication, keepalive, command and teardown phases). It then recovers framing and integrity fields such as lengths, delimiters, type bytes, checksums, MACs and counters. Finally it builds a minimal replay harness and keeps both the original capture and the reduced sequence so the result can be checked.

• Identify client and server roles, session boundaries and reset conditions before decoding field meanings.
• Keep one canonical transcript of a successful exchange.
• Change one field or one message at a time while replaying.
• Report the exact mutation that flips the server from rejected to accepted, or the reverse.

Install the skill from the ZIP or the Claude apps

Download the skill ZIP from this page and unzip it. For a personal setup, place the unzipped folder in ~/.claude/skills/. For a single project, place it in the project's .claude/skills/ folder instead.

Alternatively, upload the skill through the Claude apps: open Settings, go to Capabilities, then Skills, and upload it there.

The package contains three files: SKILL.md (the main instructions), references/custom-protocol-replay.md (the state-machine checklist, transcript checklist and evidence packaging) and agents/openai.yaml.

• Install the parent skill ctf-sandbox-orchestrator first. This skill will not work as intended without it.
• Check that the folder name and SKILL.md are where the agent expects them after unzipping.

Trigger the skill with a natural request

You do not need a special command. The skill loads when your request matches its description. Typical triggers include asking the agent to decode an unknown protocol, recover custom framing, build a replay harness, satisfy sequence or checksum rules, replay a captured session, or find the smallest message order that reaches an accepted branch.

The parent skill must already be active and must have set the sandbox assumptions. If it has not, the skill tells the agent to return to ctf-sandbox-orchestrator first.

• Decode this custom binary protocol from my capture and tell me which fields are lengths.
• Recover the framing and sequence rules in this session, then build a replay harness.
• Find the smallest set of messages that gets the server to the accepted branch.
• Work out how the checksum is computed so the replayed messages are accepted.

Supply the inputs the skill needs

The skill works on material you provide. It needs a sandboxed CTF target and captured session data. Without a capture, the agent has no transcript to recover framing or state from.

The skill's reference file asks for a few things to be kept together so the replay can be verified: message identity (type, offset, length, direction, sequence or delimiter), acceptance state (prior message dependency, negotiated value, checksum or nonce source) and the replay proof (minimal transcript, harness input and the server response or side effect).

• A sandboxed CTF target you are authorised to test.
• The captured session data, including message direction and order.
• Notes on any known session resets, handshake boundaries or duplicate messages.

Safety notes: scripts, network and authorisation

This skill runs no scripts and uses no network access, according to its listing. It is written for sandboxed competition targets. It does not establish authorisation, so it should not be pointed at live or third-party systems.

Replies are in Simplified Chinese unless you explicitly request English. Ask for English in your request if you want English output.

• Use it only against sandboxed competition targets you are authorised to test.
• Do not use it on live or third-party systems. The skill itself does not check authorisation.
• For generic packet capture analysis with no stateful replay need, the skill says to switch to the broader PCAP skill.

Tips for better results

Give the agent a clean transcript of a successful exchange. The skill's checklist asks for this first, and it makes the later field-by-field comparisons much easier to trust.

Keep messages from separate sessions apart. Mixing them into one replay model is listed as a common pitfall. Record retransmits or duplicates separately so they do not distort the protocol model.

• Start from the canonical transcript rather than a broad replay attempt.
• Change one field or one message at a time and record each result.
• Keep the original capture slices next to the replay harness inputs.