apex-review
Skill1 starsChanged 2 days agoMITView on GitHub
APEX Review: a whole-system audit of something the user ALREADY uses (a product, app, business workflow, spreadsheet or invoice system, content operation, project workspace), followed by concrete optimisation proposals and, when they agree, the upgrade itself, in any field or country. Explicit use only: use this skill ONLY when the user types /apex-review or asks for an APEX review by name (for example "use apex-review", "dùng apex review"). Never start it on your own because a request says "review", "audit" or "improve": an ordinary review request, a code review, a pull request or diff, a single file, or a small task gets an ordinary review without this skill. Words after the command steer it: "improve" for continuous improvement of a system that already works, "lean" for a compact report. To build something new from an idea, the user types /apex instead.
2 other projects use a version of this.
Adopted by
Most recently taken up by
- Asno-dev/apextheir own version, first seen Oct 15, 2026
- gvago/unofficial-qodo-review-skillstheir own version, first seen Sep 24, 2026
Other versions2 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 2
- gvago/unofficial-qodo-review-skills★ 2same name
- Asno-dev/apex★ 1same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repositoryOne other file it gives its agents.
- apexSkill
The file
We split this skill into the rules it gives an agent
10 rules. About 0 are common, 9 are distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈9
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 17
Traceability: every issue and wish reaches a final status with evidence (08).
- Distinctive: line 37
Exception: if the owner's first message already says what must stay untouched *and* tells you to decide and proceed, do not wait.
- Distinctive: line 50
All of these must be true before handover.
- Distinctive: line 60
If the owner delegates ("cứ chọn phương án bạn thấy hợp nhất và làm tiếp", "you decide", "làm luôn"), decide in one line and build in the …
- Distinctive: line 65
every slice has real tests or a real run.
- Distinctive: line 86
Verify in the target app: recalculate spreadsheets in a real engine with zero error values, and use only functions their version supports.
- Distinctive: line 91
Audit the counterparties, not only the owner: check penalties, fees and deductions claimed against the contract and legal caps, and send a …
- Distinctive: line 92
Minimise data: no ID, bank or health data copied into shared files.
- Distinctive: line 100
Close findings with a fix or a ready replacement,: not a warning.
- Distinctive: line 104
Start every substantial reply with a summary of 10 lines or fewer.
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
APEX Review: audit and optimise what already exists (v4.5)
What this skill is for
The owner brings something they rely on, often their livelihood. It is usually an incomplete or drifted attempt at what they originally wanted. Often they believe it is already as good as it gets, while it quietly loses time, money or safety every day. A strong model can review files competently. What it reliably lacks is thoroughness that survives to the end: it notices an issue and never fixes it, fixes the visible thing while the costly one stays hidden, or calls the job done without proof.
This skill adds the missing layer and leaves the model's judgment free:
- Completeness: sweep every important angle (
07). - Traceability: every issue and wish reaches a final status with evidence (
08). - Proof: parity with what already works, protected history, and tested change scenarios (
08). - Field depth for any context: a domain brief for the owner's exact field and jurisdiction (
09). - Continuous improvement: a measured baseline and a habit of getting better (
10).
The goal: the owner ends up thinking "I didn't know it could be this much better, and nothing I relied on broke." Always reply, and write user-facing deliverables, in the owner's language.
Turn 1: findings first, then the root-vision questions; no file changes
- Read the workspace for real. The quick inventory is
python "<skill_dir>/scripts/xray_scanner.py" --dir "<target>" --name "<name>" --scan-only; it writes nothing. Then open the actual files, run what is safe to run, recompute the numbers, and read the records. Build the domain brief (09) for this field and place. - Lead with what you found (usually 3–6 items), ranked by money, data and safety risk, each with evidence. Never pad; if the workspace is healthy, say so and name what could still be better: file and line, the number you computed, the record affected, and what it costs the owner. Proposing what to fix is fine. Do not edit, move or create files in the workspace yet: you do not know what the owner considers sacred.
Ask at most 3 questions, each answerable in one line, set apart from the findings:
- Original vision: "What did you originally want this to do, start to finish?"
- Pain and gaps: "Where does it waste your time, and what did you want that it never did?"
- Sacred: "Which files, data or habits must stay exactly as they are?"
Merge or drop any the workspace already answers. Stop and wait.
Exception: if the owner's first message already says what must stay untouched and tells you to decide and proceed, do not wait. Take the verified backup, then continue with the outcomes below in the same reply.
If the workspace was built or reviewed with APEX before (the inventory lists "Earlier APEX files"):
docs/apex-blueprint/ordocs/apex-review/from an older version: read it, but with care. Older templates were pre-filled, so much of it may be boilerplate that nobody decided. Trust only the parts that are clearly specific to this project (names, numbers, features the owner would recognise). Offer those back to the owner in one line to confirm, and ask about the rest. Ignore its scores, "PASS" and "approved" lines. A promise in the old ledger that the app does not keep is a finding only if the owner confirms they wanted it. Leave the old ledger in place, unedited._apex/from a current version: continue it. Keep the existing rows and update their statuses; do not start a new trace.- New process files go in
_apex/(LEAN). ForSTANDARDwork where an olderdocs/apex-review/exists, stay in_apex/or ask the owner before moving the old folder aside.
Why this order: findings prove the review is worth the owner's time. The vision questions stop you from polishing the wrong thing. A read-only Turn 1 protects data the owner has not yet called sacred.
Turn 2 onward: outcomes, not a script
All of these must be true before handover.
Gap matrix shown, short and plain:
- keep: what already serves the vision;
- built wrong: what drifted;
- never built: vision items with no implementation;
- beyond expectation: cheap, high-value additions and silent-leak fixes.
See
01and02; the archetypes there are examples, not a boundary. Every item becomes aTRACE.mdrow.- Path chosen:
- If the owner delegates ("cứ chọn phương án bạn thấy hợp nhất và làm tiếp", "you decide", "làm luôn"), decide in one line and build in the same reply.
- Otherwise, show at most 3 options with pros, cons, best-fit and your recommendation, only where the choice is genuinely theirs.
- Safe first: slice 1 is a verified backup of the current state plus locked history. Put the backup beside the workspace (for example
<workspace>_backup_<date>/or a zip) or where the owner says, on a different drive or cloud location when one is available, and tell the owner where it is. Finished records (issued invoices, signed contracts, closed periods) are protected in place: a hash, read-only where that does not disturb the system, never moved or edited. Live data the system still writes to (a running database, the current sheet) is backed up but not set read-only; change it only with a tested rollback. - Heaviest flow next: then the owner's biggest pain or volume, then the rest of never built, then the delighters.
- Proven:
- every slice has real tests or a real run;
- parity holds: existing valid inputs give the same results unless intentionally corrected, and you state it when they do;
- the number of change scenarios the context profile requires pass (08 §4; usually 3).
- Tailored, then swept: answer the context profile (07 Step 0), then sweep with the rules it activates (07); the others are
N/Awith a reason. AllTRACE.mdrows are final. - Measured: a baseline and a before/after on the few metrics the owner cares about. When the system will keep running (profile g) and already works, run the continuous-improvement loop (10) and leave
IMPROVEMENT-LOG.mdplus a re-review prompt; for a one-off review, a before/after is enough. - Handed over:
- what changed, before and after;
- the daily use;
- how to change prices or rules safely;
- backup and rollback;
- the short traceability table ("Bạn cần → Nằm ở đâu → Đã kiểm thế nào").
- No silent deferral: a High row (law or compliance, safety or health, money at risk, irreversible or historical data, the owner's root vision, their #1 pain) ends
DONE, with at least a minimal working version or exact actionable instructions, or endsDEFERREDonly withagreed: "<owner's words>". A line saying "check later" is not coverage. - Looking ahead: the domain brief's next-12–24-months section (scheduled mandates, price risk and hedging, financing, the growth path toward the owner's long-term goal) has been checked against the plan.
The daily operators decide the medium
Installing Python or packages for your own analysis, generation and testing is one command and costs nothing; use code freely. But what the owner's people operate every day must run where they already live: their phone, Excel, Outlook, Zalo or existing platform. See 08 §2.
The failure to avoid: an accountant who only uses Excel receives a Python engine, and its summary sheet shows #NAME? when opened in real Excel. Code that builds and verifies a native workbook gives both: the operators keep their tool and the owner gets reproducibility.
- Shared workflows must not depend on one machine.
- Verify in the target app: recalculate spreadsheets in a real engine with zero error values, and use only functions their version supports. If you could not, say so and give a 1-minute check.
Beyond the obvious fixes
These are easy to miss. They are generic but not universal: apply the ones the context profile activates (07 Step 0), mark the rest N/A with a one-line reason, and scale depth to the stakes:
- Audit the counterparties, not only the owner: check penalties, fees and deductions claimed against the contract and legal caps, and send a ready dispute message.
- Minimise data: no ID, bank or health data copied into shared files; set retention and deletion.
- Sanity-check the data: against itself and field norms (impossible dates, yields above the norm, stock contradictions, claims the owner's own data disproves).
- Standard instruments of the field first (09).
- Second-order effects of every fix:
- remedy past periods;
- reprice when compliance raises cost;
- tell affected people first, with ready messages;
- plan for the system's own failure (owner account, silent-stop alarm, permissions, a parallel run or drill before cut-over).
- Close findings with a fix or a ready replacement, not a warning.
Respect the owner's time
- Start every substantial reply with a summary of 10 lines or fewer. Put details in files written in the owner's language.
- Keep replies short. After the summary, include only what the owner must read to decide or act: the findings, what changed, what they do next. Aim for a reply the owner can read in about two minutes. Do not repeat in chat what a file already says; link the file.
- Write only the documents someone will use. One how-to for the operators beats several overlapping guides. Do not produce documents that restate the chat or exist to show effort.
- Leave no clutter: process files (trace, ledger, notes) go in one
_apex/folder (ordocs/apex-review/in STANDARD/DEEP mode), never loose in the owner's working folder. User-facing files are few, plain, and in the owner's language. - Lenses, gates, zone colours, scores and phase names are internal; do not show them unless asked.
- For rules, rates and fees, give the identifier and date, and mark unverified items with who can confirm them.
Rigor modes (choose silently)
| Mode | When | Shape |
LEAN (default) | Small workspace, one to a few operators | _apex/TRACE.md and _apex/STATE.json, plus one compact review (_apex/REVIEW.md, ≤ 120 lines) if useful. |
STANDARD | Multi-module app or business-critical workflow | Ledger in docs/apex-review/; gates per slice. |
DEEP | The owner asks, or the stakes are high (money movement, health, legal, irreversible data, safety) | STANDARD plus one synchronous critic-subagent pass, run once; never sleep-poll. |
In chat-only hosts, keep TRACE.md and the change scenarios as tables in the reply, and replace script gates with a traced simulation: 1 normal case and 2 failure cases.
Which references to load
Load a reference only when its moment comes. This file and references 07–09 take precedence over 01–06 wherever they differ.
| Moment | Load |
| Turn 1, before the first reply | 09 (domain brief) |
| Building the gap matrix; a genuine fork to put to the owner | 01 |
| Before presenting the plan and before handover | 07 (context profile, lenses), 08 (trace, definition of done, change scenarios) |
| Critique at medium or high stakes | 03 |
| Planning the upgrade architecture | 04 |
| Starting to change files | 06 |
| The system already works and the owner wants it better | 10 |
STANDARD / DEEP ledger work | 05; 02 for worked examples of field depth |
A small LEAN review needs only 09, 07, 08, and 06 once you change files.
Scripts (CLI/IDE hosts)
- Turn-1 inventory (read-only):
python "<skill_dir>/scripts/xray_scanner.py" --dir "<target>" --name "<name>" --scan-only - Scaffold (Turn 2+):
LEAN: the same command with--leaninstead of--scan-onlycreates_apex/TRACE.mdand_apex/STATE.json.STANDARD/DEEP: without either flag it createsdocs/apex-review/.--domainacceptsDOM-01..07or any free-text field.
- Gate audit:
python "<skill_dir>/scripts/review_gate_audit.py" --dir "<target>" --gate <1..6> --scan-target <files or folders you built or changed>. Point the stub scan at your own work, not the whole app: aTODOalready in the owner's code is a finding to report, not a reason to edit code nobody asked you to touch. It works in both modes: with a ledger it checks the documents; with only_apex/TRACE.mdit runs the LEAN checks. - What the gates check (details in
06):- Documents are present, substantive and free of
[[FILL: hint]]markers. The check is structural, so write them in the owner's language. TRACE.mdrows are checked at every gate once the file exists:DONEhas evidence, and a High row isDEFERREDonly withagreed: "<owner's words>". Keep the status and weight codes in English.- Gate 3 (ledger modes): a critique score ≥ 85 for the plan once the flaws the critique found are fixed, read from the
TOTAL UPGRADED SCOREline or fromupgraded_health_scorein STATE. - Gate 5:
--scan-target, and at least one proven slice (ledger modes; slice 1 is the backup). - Gate 6: no
OPENrows, all slices proven (ledger modes), andchange_scenarios_requiredpassed scenarios (default 3; fewer needs achange_scenarios_waiverreason).
- Documents are present, substantive and free of
- The script checks structure and evidence, not quality. Never pre-select an option or write a score, proof or status you did not obtain. Report results as plain outcomes, not gate numbers.
How the user invokes it
/apex-review <path or description> is the one command, and the skill runs only when the user types it or asks for an APEX review by name. It is a deliberate, heavier process; a request that merely says "review" or "audit" gets an ordinary review. Words in the request steer the mode:
- "improve": the system already works; run the continuous-improvement loop (
10). - "lean":
LEANmode; findings plus an upgrade in one compact report. - "organise" or "clean up": findings on organisation and risk, then a cleaner structure with protected history (
04§2). - "delta": a before/after comparison on the baseline metrics (
10), plusreview_gate_audit.pyto list what is still open.
This is a deeper job than an ordinary review. If the user only wants a code review, a pull request or a single file looked at, say that /apex-review is the whole-system audit and do the plain review they asked for instead. If they want something new built from an idea, that is /apex.
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
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.

