Guide · Agent skill
How to use the Pentest Project Blackboard agent skill
Updated · written with AI from the project's own documentation and checked against our catalogue data
The pentest-blackboard skill is a set of instructions that teaches an AI agent to write each confirmed finding, attack path and piece of evidence into a SQLite-backed project blackboard while it works, rather than relying on the chat history. It is written for the CyberStrikeAI penetration-testing product and expects a project to be bound to the conversation first. After this guide you will know how to install the skill, bind a project, trigger the behaviour with plain requests and check that facts are being stored.
Pentest Project BlackboardAIPentest/CyberStrikeAI · Apache-2.0 Download the skill Quick start
- 1Download the ZIP. Download the skill ZIP from this directory page. It contains SKILL.md.
- 2Unzip into a skills folder. Place the skill in ~/.claude/skills/ for personal use, or in .claude/skills/ inside a project for that project only.
- 3Or upload in the Claude apps. If you use the Claude apps, go to Settings → Capabilities → Skills and upload the ZIP.
- 4Bind a project in CyberStrikeAI. Bind the conversation to the project before testing, because the blackboard tools fail without a bound project.
- 5Ask for recorded results. Give a scoped, authorized request and ask the agent to save confirmed facts and vulnerabilities to the project, then check the stored keys and summaries.
What the skill makes the agent do
Without a shared record, an agent's findings live only in the chat, and long sessions or context compression can drop details. This skill changes that by having the agent store facts in the project blackboard as it goes. Each fact has a key, a one-line summary and a fuller body with the reproducible context.
Later in the session the agent can look up stored facts by key, search them, and file formal vulnerability records. The source describes the separation as Fact for environment and exploit context, and record_vulnerability for deliverable findings, with both allowed for the same issue.
- Writes facts with upsert_project_fact; writing the same fact_key again overwrites the earlier entry.
- Reads full detail with get_project_fact, and lists or searches facts with list_project_facts and search_project_facts.
- Files deliverable findings with record_vulnerability and checks for duplicates with list_vulnerabilities first.
- Marks false positives with deprecate_project_fact and restores them with restore_project_fact.
- Links related facts with confidence levels of confirmed, tentative or deprecated.
Install the skill from the ZIP
Download the ZIP for the skill from this directory page. The package contains one file, SKILL.md. Unzip it so that the skill folder sits in one of two places: your personal skills folder at ~/.claude/skills/, or a project-specific .claude/skills/ folder inside the project you work in. Use the personal folder if you want the skill available everywhere, and the project folder if you want it only for that codebase or engagement.
If you use the Claude apps instead of a folder, upload the ZIP through Settings → Capabilities → Skills. Either way, the agent should pick up the skill from its name and description, so you do not need to call it by a special command.
- Personal install: ~/.claude/skills/
- Project install: .claude/skills/ inside the project folder
- Claude apps: Settings → Capabilities → Skills, then upload the ZIP
Bind a project before you start
The skill stores state in the project blackboard, which is tied to the current conversation. The source is explicit that the conversation must already be bound to a project, and that without one the tools return an error. Bind the project in CyberStrikeAI before you give the agent any test work.
If no project is bound, the agent should say it cannot write to the blackboard and keep a summary of the evidence in the current turn. Check that the binding exists before you start a multi-step test, not after the session is half done.
- Confirm the conversation is bound to the correct project.
- Confirm the testing scope is authorized and written down.
- Ask the agent to search existing facts first if the project already has history.
Trigger it with natural requests
You do not need special syntax. Describe the work in plain terms and mention that results should be saved to the project, so the agent uses the blackboard rather than only replying in chat. The skill's description mentions managing project facts, the blackboard index, writing evidence and avoiding context loss after compression, so requests on those topics are the most likely to trigger it.
Good requests name the target, the goal and the record you expect. Ask the agent to report back what it stored, so you can check the keys and summaries.
- Record the open ports and service versions you confirm on the staging host as facts under target/.
- Save the working login path you found as an auth/ fact, and note whether the evidence is confirmed or tentative.
- Search the blackboard for any finding about the login form before you retest it.
- When the SQL injection on the login page is verified with a reproducible request, file it as a vulnerability record and link it to the fact.
- Summarise what is stored in the blackboard for this project and list any facts still marked tentative.
Check how facts are named and stored
Keys use lowercase category/slug form, such as target/primary_domain or finding/sqli-login. The allowed environment categories are target, auth, infra and business, and the discovery categories are finding, chain, exploit and poc, with note also available. The source tells the agent to keep one key per discovery rather than splitting it, so the context stays together.
A summary should say what the issue is, where it is and how it is triggered or verified. The body holds the steps, raw requests or responses, observed behaviour and evidence. Relationship links are required for finding, chain, exploit and poc entries, typically a discovered_on link from the target fact.
- Good summary: login form at /login is vulnerable to SQL injection, confirmed with the quote-based request in the body.
- Weak summary: SQLi exists. The source says summaries like this are not acceptable.
- Use confidence confirmed only when evidence pins the issue down; use tentative for leads still being checked.
target/primary_domainSafety notes: authorization, scripts and network
The skill drives active exploitation, credential reuse across services and proxy switching. The source is clear that using these techniques against systems you do not own or have permission to test is illegal. Do not run this skill against any target until you have written authorization that covers the scope.
The directory metadata for this skill lists no bundled scripts and no network use. The skill text does refer to other tools and other skills, such as terminal curl, browser navigation and the component-vuln-intel, proxy-tool-bootstrap and pentest-verification skills. Those are not included in this package, and the instructions depend on the CyberStrikeAI product's tools being available.
- Keep credentials and secrets out of the summaries; the source suggests redacting and following the project's secret-handling policy.
- Record every write-type change in the reply and, where useful, as a note/change- fact with the host, type, location and rollback command.
- Expect failures outside CyberStrikeAI, because the referenced fact and vulnerability tools exist only in that product.
Tips for long sessions and multi-agent work
Store facts as you confirm them rather than at the end of the session, because context compression can drop details that were never written down. When an agent gets blocked, the source tells it to search existing facts before trying again, which avoids repeating dead ends.
In a coordinated setup, the lead agent should write sub-agent findings to the blackboard. If a sub-agent has no database tools, ask it to end with a list of items to store, including a suggested fact_key, summary and body or proof-of-concept points.
- Read the auto-injected index first, and call get_project_fact when the summary is not enough.
- Keep one fact_key per finding and overwrite it, instead of creating many near-duplicate keys.
- Omitting links on an update keeps existing links; supplying links replaces all incoming links, so be deliberate.
What people use it for
- Logging a multi-stage web application test so each confirmed endpoint, login path and finding is stored under its own key.
- Coordinating several sub-agents on one engagement, with the lead agent writing their findings into a single shared record.
- Keeping per-target evidence linked, so a finding can be traced back to the target fact where it was discovered.
- Recovering context after a long session or compression by searching stored facts instead of rereading the chat.
Frequently asked questions
How do I install the pentest-blackboard skill?
Download the ZIP from this directory page and unzip it into ~/.claude/skills/ for personal use, or into .claude/skills/ inside a project. In the Claude apps you can instead upload the ZIP under Settings → Capabilities → Skills.
What do I need to use pentest-blackboard?
You need the CyberStrikeAI product with its project blackboard and fact and vulnerability tools, a project bound to the current conversation, and an authorized penetration-testing scope. The skill also refers to other skills that are not included in this package.
Is pentest-blackboard free to use?
The directory lists the licence as Apache-2.0. Check the licence file in the source repository for the exact terms before you redistribute or modify it.
Why does the skill say my project tools failed?
The blackboard tools require the current conversation to be bound to a project. Bind the project in CyberStrikeAI first, then retry the request.
Can I use this skill outside CyberStrikeAI?
The source says no in practice: the referenced tools upsert_project_fact and record_vulnerability exist only in CyberStrikeAI, so the instructions will fail elsewhere. The source does not name alternative tools, so you would need to find another approach for your own environment.
Does this skill run scripts or use the network?
The directory metadata lists no bundled scripts and no network use. The skill's instructions do refer to tools such as terminal curl and browser navigation, which come from the CyberStrikeAI product rather than from the skill files.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
This skill suits security testers and red-team leads who already work inside CyberStrikeAI, run multi-step authorized tests, or coordinate several sub-agents and want one structured, linked record of findings and evidence. It is a poor fit for anyone without written authorization for the targets, for anyone using a different tool, or for general coding and writing work.
Download the skill