Technique Ctf

Technique CTF

Solve any flag-style challenge with a category-agnostic loop: triage the artifact, identify the exact spec or format, drive the smallest correct interaction, and let the target's own responses steer each next input — then verify and submit the flag.

When this skill applies

• Any controlled lab/CTF objective: a remote service, a downloadable bundle, a puzzle artifact, or a skeleton client to complete.
• Before reaching for a domain skill, to decide which *-ctf skill(s) to load and how to structure the attack.
• Whenever a target emits parseable feedback — validation errors, diffs, status codes, timing, or partial output — that can be used as an oracle.
• When several independent challenges (or independent parameter variants of one challenge) can be worked in parallel.

Operating model

1. Triage. Inventory what you were given: hash and list files, read the brief literally, note the target host:port, the stated flag format, and every concrete value (IDs, keys, payload strings, ports). Identify the dominant artifact class — it selects the domain skill.
2. Identify the spec. Name the exact format/protocol/algorithm before writing an exploit. Match magic bytes, headers, wording, and cited standards. The brief usually states the spec and the exact expected input — build to that, not to placeholder values in any example code.
3. Minimal correct interaction. Send the smallest well-formed unit and observe. Probe liveness both ways: some services greet with a banner, others reply only after you send — send-then-recv, do not hang waiting for a banner.
4. Read the oracle. Treat every response as signal. Precise validators say exactly what is wrong (expected X but got Y, invalid state, a stack trace, a length, a 401 vs 403, a timing delta). Walk the error chain field by field; each rejection names the next thing to fix. This beats guessing and beats blind brute force.
5. Iterate / brute the smallest unknown. When one field is genuinely ambiguous after the oracle is exhausted, brute only that minimal space (e.g. a counter × a type bit), reconnecting/resetting state per attempt. Never brute a large space blindly when a signal exists.
6. Extract and verify. Pull the flag with a format regex, sanity-check it, and only then submit. If a submission API exists, use it; confirm acceptance. Do not paste flags into external services.

Cognitive stance (negative constraints)

• Evidence first. Do not claim a solve, a working exploit, or a root cause without a reproduced observation. If a step was skipped or failed, say so.
• Oracle over guessing. Exhaust the target's feedback before brute force; exhaust a small brute before a large one.
• Minimal action. Prefer offline decode of a supplied artifact over live interaction; prefer one crafted request over a scanner. Loud, broad tooling wastes time and corrupts stateful targets.
• No rabbit holes. Time-box a dead end. If the oracle stops yielding, re-triage: wrong layer, wrong spec, or an assumption to test — not more of the same input.
• Verify the flag. A plausible-looking string is not a flag until it matches the stated format and (if possible) is accepted.

Category routing

Load the domain skill that matches the dominant artifact; load more than one only for genuinely cross-domain bundles.

Parallelization

Independent work should run concurrently, not serially.

• Dispatch one subagent per independent challenge, or per independent variant of a stubborn challenge (different spec guess, different parameter branch). Give each the concrete target, the exact spec, and any reusable code so it does not rediscover the basics.
• Split a single challenge only along a true seam (recon vs. exploit dev, or independent sub-flags) with no shared mutable state.
• Keep results, not transcripts: each subagent returns the flag and the minimal working method; you retain the conclusion.

Tool routing

• Remote loop: pwntools (remote, recvuntil, sendline), or raw socket for exact byte control. Script every interaction — never solve a stateful service by hand.
• Transforms: CyberChef, xxd/hexdump, base64/base32, jq, Python struct/construct for framed formats, binwalk/file/strings for unknown blobs.
• Recon of the artifact, not the network: read the files first. Reserve scanners for when the brief points at a broad service surface, and keep them scoped.
• Reproducibility: keep a single script per challenge that goes from connect to flag, so a fix is one edit and a re-run.

Flag discipline

• Match the stated format; default regexes: HTB\{[^}]+\}, flag\{[^}]+\}, [A-Za-z0-9_]+\{[^}]+\}, or the exact prefix the platform uses.
• Verify before submitting; if the platform has a submit API/tool, use it and confirm the acceptance message.
• Do not exfiltrate flags, source, or target data to external/third-party services; keep everything in the lab context.

Quick pivots

• Encrypted HTB bundle: use hacktheblue for Sherlock archives and hackthebox for other challenge archives. Prefer the known platform password over brute forcing or trusting a downloader's stale default.
• Timeout on connect: the service wants input first — send the minimal frame/request, then read.
• "Invalid" on a unit you believe is valid: an encoding assumption is wrong (raw byte vs. ASCII string, endianness, off-by-one length, CRC seed). Try the alternate encoding before rebuilding everything.
• Errors reference a counter/sequence/state: the target is a state machine — advance every counter it tracks in lockstep; reconnect to reset when you lose sync.
• Same error regardless of input: you are at the wrong layer or failing an earlier gate (auth, framing, checksum) before your payload is even parsed.
• Big search space: find the oracle that collapses it (an error that leaks length/position, a timing side channel, a partial match) instead of brute forcing.
• Stuck: re-read the brief for an unused concrete value; it is almost always load-bearing.
• Total silence is a different failure from a rejection. A target that never answers may be broken or not listening, not wrong-input. Before iterating on payloads, prove the far side is alive and consuming: send a deliberately malformed unit and check you still get some reaction, verify the peer actually drains what you send (a queue that accepts data proves buffering, not processing), and plant a canary that needs the minimum machinery to echo. Endless silence with zero diagnostics usually means the environment, not the exploit.
• Verify the instance is yours before trusting it. A platform that allows only one live instance at a time will silently replace or expire the previous one while its status still reads "ready", so a cached ip:port can point at a dead or foreign container. Reconnect and match the banner before concluding a technique failed, and serialize work when instances are exclusive.
• Re-derive a parameter per instance when the target regenerates it. If a value changes between spawns (key, IV, nonce, TLE, port), any artifact captured from an earlier instance is poison — collect every dependent input from the same live instance.

Resources

• references/harness-and-oracle-patterns.md — reusable pwntools/socket remote-loop templates, an oracle-error-walking pattern, a minimal-matrix brute template that reconnects per attempt, flag-regex helpers, and a subagent parallelization pattern for independent challenges. Load when building the solve script for a remote or stateful challenge.