agentleFS
moiz-73moiz-73/glassbox/skills/recon/SKILL.md

recon

Skill0 starsChanged 44 days agoMITView on GitHub

Establish ground truth about a system you are authorized to work with. Use when documentation and observed behavior disagree, when mapping where logic lives, classifying protection posture as none, silent, or interactive, mapping quotas to the axis that moves them, auditing utilization before adding capacity, or probing resources before parallel work. Produces a dated facts table where every load-bearing claim has a live capture. Not for protocol decoding (`protocol-cartography`) or retry, lease, and run state design (`flow-modeling`).

32 other projects use a version of this.

Adopted by

Most recently taken up by

Used with

Projects with it also use reporting, osint and 3 more.

Other files people use alongside this one.

Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
  • GitHubSkill

    reporting

    Report formatting guidelines for gh-aw workflows (header levels, collapsible details, run-ID links)

    Also in 3 projects that use this

  • JosBen74Skill

    osint

    OSINT-sökning och informationsinhämtning från öppna källor för Länsstyrelsen VG

    Also in 4 projects that use this

  • samklee1Skill

    web-pentest

    Orchestrates authorized web app pentesting on Kali Linux using Opus sub-agents with difficulty-aware planning (TDA/EGATS), browser-first exploration, iterative verification, two-layer review (per-sub-agent + phase-level), optional cross-model gap audit (GPT/codex via codex-plugin-cc), and enforced KB updates. Trigger on: pentesting, penetration testing, web app security assessment, vulnerability assessment, OWASP testing, PTES, attack surface mapping, API security testing, pentest reporting, or 'help me test this app'. NOT for network-only pentesting, malware analysis, reverse engineering, or CTFs.

    Also in 3 projects that use this

  • SentrySkill

    create-pr

    Create pull requests following Sentry conventions. Use when opening PRs, writing PR descriptions, or preparing changes for review. Follows Sentry's code review guidelines.

    Also in 3 projects that use this

  • NVIDIASkill

    research

    Before designing a feature, research how comparable libraries solve the problem, check best practices, known pitfalls, and upstream changes. Use when starting a new feature, investigating an unfamiliar area, or the user asks to research something.

    Also in 3 projects that use this

The file

We split this skill into the rules it gives an agent

7 rules. 0 are common, 7 are distinctive, and 0 are very specific to this project.

  • 0

    Common

    Most of the 27 edited versions we read keep these.

  • 7

    Distinctive

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Classify before you automate: Every surface gets a protection class: none, silent, or interactive.

    line 49
  2. Distinctive:

    Numbers decay: Every measurement carries a date and an evidence grade.

    line 54
  3. Distinctive:

    Write one paragraph, in a file, before touching anything technical: what the system is, the authorization basis (your own system, a written …

    line 71
  4. Distinctive:

    Every later judgment — "may I automate this step?", "may I push through this friction?" — is a scope question wearing a technical costume.

    line 75
  5. Distinctive:

    Before adding capacity, audit utilization of what you already have: usage counters, served versus idle capacity, headroom.

    line 141
  6. Distinctive:

    Update it only from evidence after a run — never mid-run from a hunch.

    line 150
  7. Distinctive:

    Before any parallel run, measure the machine live: CPU, free memory, disk pressure, and any other hard limit.

    line 156

“Common” is measured: kept by at least 6 in 10 of the 27 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Recon

Recon is the perception discipline: before any model, harness, or automation depends on a system, establish what actually happens there. Its single product is a dated facts table where every load-bearing claim is observed live and carries the capture or probe that produced it. Every later artifact — a protocol contract, a state machine, a retry policy — is only as reliable as this table.

Decision table

SituationActionLoad
Starting work on an unfamiliar systemWrite the authorization scope statement; keep it with the workhonest-fences
Docs, a client library, and live behavior disagreeRank the sources; reproduce the load-bearing claim liveMethod 2
You cannot tell where logic or friction livesMap surfaces; mark the decisive one and deprioritize the restreferences/surfaces-and-fences.md
Deciding whether a surface can be automatedClassify protection posture honestly before automatingMethod 4
Limits or entitlements shape the strategyMap each limit to the axis that actually moves itMethod 6
Tempted to add capacityAudit utilization of existing capacity firstMethod 6
About to run parallel workersProbe machine resources; derive and log the capMethod 8
A multi-step flow keeps being re-decided mid-runFreeze the branch points into a zero-think tableMethod 7
Ready to leave perceptionExit gate: every load-bearing fact is [OBS] with an artifactevidence-protocol

Golden rules

  1. The wire is the truth. Live observed behavior outranks the system's own client, which outranks documentation, which outranks memory. Any argument that starts with "surely it behaves like that" is settled by one probe, not by discussion.
  2. One surface is never the whole system. Client-rendered applications hide their logic in bundles and across origins, so "nothing on the first surface" means "look at the next surface", never "there is nothing here".
  3. Classify before you automate. Every surface gets a protection class: none, silent, or interactive. Quiet mechanics are automated within their stated limits; interactive human-verification steps are routed to a human as a designed part of the workflow. An unclassified boundary is discovered after the build, when it is most expensive and most tempting to mishandle.
  4. Numbers decay. Every measurement carries a date and an evidence grade. An undated number is a rumor, and volatile numbers must be re-verified before they influence a decision.
  5. Measure the machine before you parallelize. Concurrency is a measured quantity, not a wish; a fanout that exceeds the measured cap converts progress into contention.
  6. Stop when the system defends itself. Persistent turning-away, suppression, or tightening limits is a signal to stop, cool down, and probe once — not a defect to push through. Waiting is the correct move.
  7. The scope statement comes first. Availability is not permission; when the authorization basis is unclear, stop and ask before collecting anything.

The method

1. Scope the engagement

Write one paragraph, in a file, before touching anything technical: what the system is, the authorization basis (your own system, a written engagement, a sanctioned exercise, or an integration you were asked to build), what is in scope, what is explicitly out of scope, and where a human must act or approve. Every later judgment — "may I automate this step?", "may I push through this friction?" — is a scope question wearing a technical costume. With the scope written and dated, those judgments take seconds.

2. Rank the sources of truth

Establish the hierarchy and never let a lower rank overrule a higher one:

RankSourceTrust posture
1Live observed behaviorReproduce before relying on it
2The system's own client, SDK, or CLIA convenience layer; may lag
3Official documentationFrequently stale; treat as hypothesis
4Memory and pattern-matching from other systemsHypothesis only

Tag each fact [OBS] observed live, [REP] reported or configured secondhand, or [INF] inferred and untested, with a date. A [REP] or [INF] fact is re-verified before it influences a build decision.

3. Map the surfaces

Enumerate the surfaces before concluding anything: static content, the client-rendered application, background request traffic, native or desktop clients, command-line tools, and embedded third-party frames. For each surface, record where its logic lives, how you observe it, and whether it matters for this job. Then name the decisive surface and say why the others were deprioritized — the map is a decision document, not an inventory.

4. Classify protection posture

Probe each surface harmlessly and classify what you find:

  • none — no friction observed in the paths you are authorized to use.
  • silent — quiet mechanics that shape behavior: throttle windows, caps, entitlements, reset schedules. Automate these within stated limits, with budgets and backoff.
  • interactive — a step that explicitly requires a human. Route it to a human as a designed part of the workflow.

Record the classification as a dated observation and re-check before scaling, because postures change. The classification is a policy input, not a description: an interactive step ends the automation question, and the honest answer is a human handoff, never engineering around the wall.

5. Capture complete exchanges

For every operation in scope, capture the full exchange: request, response, timing, ordering, identifiers, and the headers that matter. Prefer the system's own client as the capture vehicle — it knows the current contract. Capture enough per operation to replay it later from a clean state. Protocol decoding proper belongs to protocol-cartography; recon's job is to know which exchanges exist and to keep the raw material.

6. Map quota surfaces and audit utilization

Vary one axis at a time — project, connection pool, address, concurrency — and observe which change actually moves the limit. Read limit headers and error bodies; record the window length and the reset epoch. Compare the shared-pool case with the per-unit case using the resources you already own. Then draw the quota model: axis to limit to window to reset.

Mapping exists to respect limits and plan capacity honestly, not to extend them. Never add identities, accounts, or keys to move an axis where the platform's terms prohibit it; when the basis is unclear, that question belongs to honest-fences.

Before adding capacity, audit utilization of what you already have: usage counters, served versus idle capacity, headroom. If most capacity sits idle, the fix is routing and wiring, not acquisition. Acquisition before an audit is spending used as a substitute for measurement.

7. Freeze branch points into a zero-think table

As you observe repeated branch points, write them into a table: situation, exact action, exact command. Read it once, then act without deliberation. Update it only from evidence after a run — never mid-run from a hunch. Re-deciding settled questions under time pressure is where consistency and cooling discipline are lost.

8. Probe machine resources before fanout

Before any parallel run, measure the machine live: CPU, free memory, disk pressure, and any other hard limit. Derive a safe cap from the measurement, keep an explicit safety buffer free, and cap every fanout at that number. When work exceeds the cap, run waves. Log the probe output and the derived cap with the run. The swarm-orchestration skill carries the full fanout doctrine.

9. Inventory leases and sessions

Anything with a lifetime — sessions, sandboxes, tunnels, machines — gets inventoried during recon: the lease object, its identifier, its affinity, its time-to-live, and its renewal mechanism. Probe survival explicitly: restart mid-job and record what rehydrates and what is lost. The lifecycle design is flow-modeling's job; recon owes it an honest inventory.

Exit gate. A facts table where every load-bearing claim is [OBS] with a capture or a probe command attached, protection classes recorded per surface with dates, and the decisive surface named. No guessing survives this gate.

When NOT to use

  • You already have a current, replayable contract for the system — build.
  • The task is decoding requests, streams, or dialects — use protocol-cartography, which consumes recon's captures.
  • The task is state machines, retries, leases, or run state — use flow-modeling, which consumes recon's facts.
  • The authorization basis is unclear, or the path requires crossing an interactive verification wall: stop and hand back to a human. Availability is not permission; a silent protection is not consent.

References

  • references/surfaces-and-fences.md — the long-form field manual for surface mapping, protection classification, quota-surface mapping, resource probes, utilization audits, and lease inventories. Read it while running a perception pass on a system you have not mapped before.
  • deep-reverse-engineering — the parent loop this skill feeds.
  • evidence-protocol — how to turn recon's facts into verifiable claims.
  • honest-fences — the authorization and boundary doctrine.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.