survey
Skill1 starsChanged 23 days agoMIT-0View on GitHub
Survey a part of a codebase the Tesla Valve way: a deep architecture review that follows each stream from spring to goal and reports where the flow is not natural, such as second channels, layers that only carry, streams that touch outside a contract, moving parts, uneven quality, and docs that say what the code does not do. It reads and reports, and changes nothing. Run it only when the user asks for it, on a topic the user names: it reads far beyond any one task.
23 other projects use a version of this.
Adopted by
Most recently taken up by
- quickchatai/quickreviewtheir own version, first seen Nov 15, 2026
- genkovich/sddtheir own version, first seen Oct 25, 2026
- PKU-DAIR/AutoScitheir own version, first seen Oct 25, 2026
- wheslleyrimar/BuildCLI-Agentstheir own version, first seen Oct 22, 2026
- LazyerIJ/self-clonetheir own version, first seen Oct 17, 2026
- steven112163/dotharnesstheir own version, first seen Oct 15, 2026
Used with
Projects with it also use daily-arxiv, exp-eval and 10 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.
LambenthanSkilldaily-arxiv
每日从 arXiv 拉取新论文,过滤相关性,auto-ingest 高优先级论文,检测 SOTA 更新
Also in 5 projects that use this
LambenthanSkillexp-eval
实验判决门:Review LLM 独立评判实验结果 → 4 种判决路径 → 自动更新 claims confidence、ideas status、graph edges
Also in 5 projects that use this
skyllwtSkillexp-pilot-eval
Pilot result evaluation — read pilot results, apply success criteria, update idea page (pilot_result, failure_reason if failed), generate PILOT_VERDICT_REPORT. Called by /ideate Phase 5 after /exp-pilot-run.
Also in 4 projects that use this
skyllwtSkillexp-pilot-run
Pilot experiment execution — read Pilot Spec YAML, write pilot code, run experiment(Confirm with the user before operation and require the applicant to conduct manual inspection), return results. Called by /ideate Phase 5. Does NOT modify wiki pages or judge pass/fail.
Also in 4 projects that use this
phanghonghaoSkillposter
HTML/CSS A4 海报生成。从沙龙/活动 Markdown 文件生成科技风 A4 海报 PDF/PNG。Use when user mentions: poster, 海报, 海报生成, make poster, generate poster, 活动海报, 沙龙海报.
Also in 4 projects that use this
- +6 more from Lambenthan
- +1 more from skyllwt
Other versions23 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 23
- skyllwt/OmegaWiki★ 1.7ksame name
- DAIR-AutoResearch/AutoSci★ 1.7ksame name
- PKU-DAIR/AutoSci★ 1.7ksame name
- skyllwt/AutoSci★ 1.7ksame name
- genkovich/sdd★ 187same name
- SNL-UCSB/literature-survey-skill★ 93same name
- Lambenthan/empiricalwiki★ 85same name
- spacedock-dev/spacedock★ 85same name
- V-Songbird/foreman★ 35same name
- jmylchreest/aide★ 19same name
13 more not listed
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 repository4 other files it gives its agents.
- AGENTS.mdAGENTS.md
- prepareSkill
- reshapeSkill
- tesla-valveSkill
The file
We split this skill into the rules it gives an agent
2 rules. 0 are common, 2 are distinctive, and 0 are very specific to this project.
0
Common
Most of the 18 edited versions we read keep these.
2
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
- Distinctive: line 66
Read three things, in this order.
- Distinctive: line 83
Every dimension asks the same question from another side: is the right path the path of least resistance here, and does the wrong path work …
“Common” is measured: kept by at least 6 in 10 of the 18 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.
Files
Survey
The tesla-valve skill looks at the code one task touches, and the reshape skill at the checks in one scope. Neither asks whether the landscape still flows after fifty tasks have each dug where they stood. This skill asks that. It reads and it reports. It changes no code, no config and no document.
It holds no rules of its own: every finding follows from the tesla-valve skill. Read it first, and its reference for each language in the scope. What this skill adds is the method, and the same view from further away: module depth, coupling and scale are not other subjects, they are what a stream looks like from above.
1. Topic and focus
The user names a topic and, after it, a focus:
checkout, where the delivery date is decided.
- The topic says where to look: a module, a subsystem, or a stream such as "how a price reaches the invoice". Resolve it to paths.
- The focus says what the user wants to know. It orders the survey and does not blind it: a severe finding outside the focus is still reported. Without a focus, every dimension in step 4 counts the same.
- Without a topic, ask for one before you gather anything, and offer the two or three areas that changed most in the recent history as candidates: a fault there is paid again with every change. A survey of a topic nobody named spends a deep review on a guess.
With a topic, tell the user in one line which paths and which question you took from it before you read further, and do not wait for an answer.
2. Gather cheap, judge here
Most of a survey is gathering: which files, who calls what, where a name is spelled, what CI runs. The judging is the smaller part, and the one that needs the strongest model.
- Where the host has subagents, give them the gathering, on a cheaper model where the host lets you choose one: one bounded question each, with a cap on the length of the answer. The map, the findings and the ranking stay with you. A topic you can read in one sitting needs no helpers: a brief longer than the reading it saves is load.
- A gathered answer is a claim. Open the lines it cites before a finding rests on them.
- Nothing is written into the project, by you or by anyone you hand work to. A helper does not know that unless its brief says so: say it in every brief, and give the helper read-only tools where the host has them. Notes you keep on the way go into the system's temporary directory.
3. Map
Read three things, in this order.
- What the project says about itself: its architecture docs, its decision records, the instructions it gives its agents. These are claims the survey tests, and decisions it does not reopen without a cost to show.
- What holds: the enforcers in its config and its CI. A rule that nothing enforces is a convention, whatever the docs call it.
- The streams: map each stream of the topic the way the tesla-valve skill maps a feature: spring, check, refusal, confluence, goal. Add what a task never needs to write down: every layer the stream crosses on its way, and every place that keeps a copy of its fact.
Findings attach to streams, not to files. "The postcode is checked in three
places" is a finding. "checkout.rs is long" is not.
4. Read the map along each dimension
Every dimension asks the same question from another side: is the right path the path of least resistance here, and does the wrong path work against itself? Each is a judgment call, so each carries its why and one example.
The dimensions are ways to look, not a form to fill in. Most of a healthy topic is clean along most of them, and a finding kept so that a dimension has one is load. A dimension with nothing to report is left out of the report.
- Springs. Does each stream from outside have one check, where it enters,
and does every refusal go somewhere that says why? A value used before it is
parsed can be anything. A value checked again downstream has two checks that
drift apart. A refusal turned into a default hides the bad data from the one
who could fix it: an unknown stock status that becomes
in_stock. - Second channels. Is each fact decided in one place? Two sources of truth
drift, and no type catches it:
delivery_datereads the warehouse itself whilestock_ofalready does. - Depth. How many layers does a stream cross, and how many of them change
it? A layer that only carries is one more place every reader visits and
every change touches. Delete it in thought: if nothing has to be written
anywhere else, it carried nothing. A
StockServicethat forwards each call toStockRepositoryis such a layer. The opposite counts too: one function where five streams meet and nobody can name the confluence. - Coupling. Streams meet at a confluence, through a contract. Where else do they touch? A change to one stream that forces edits in modules that never use it will be made wrong or not at all. Look for a module that reaches into the fields of another, a mutable value two streams share, and a context that carries ten values to a function that reads one. And look at the direction: water flows one way. The postcode parser that imports the checkout page to word its error flows back upstream, and neither can change alone.
- Moving parts. What is only right if someone remembers it, chooses it or
times it? A flag that picks the behaviour, an
init()that must run first, a list of handlers every new handler must be added to by hand. - Stale facts. Where is a fact about the world read in one place and acted on in another? Between the two the world moves: stock is read when the page loads and decremented, unconditionally, when the order is placed.
- Volume. What happens to the channel when the stream carries a thousand times more, or when the tenth variant arrives? Find where it narrows: a whole table read into memory, one round trip per order line, one lock around every order. Count the places a new variant must edit. Ask this only for growth that a feature or the project's own goals name: a channel dug for water that never comes is load.
- Even ground. Is the same thing done the same way across the scope? Each further way is a convention the reader must learn, and a place where the stricter rule does not reach: orders parse into types and returns pass strings, or a lint is on in one package and off in the next. Name the difference, and which way the project's enforcers favour.
- Said and built. Where does a doc, a decision record, a comment or a name
say what the code does not do? The words are a second channel, and the next
change will trust them:
// validated upstreamabove a value that nothing upstream validates. - Tests. Which intent has no test, above all across parts: a protocol, atomicity, recovery from a failure? Which check at a spring is never fed bad data? Which tests pin only a shape the compiler already holds?
- Load. What does no feature need? A guard for a case that cannot occur, an option nobody sets, an interface with one implementation and no second in sight, a channel no stream flows through any more.
5. Verify
A finding is a claim until it is checked. Each one that reaches the report:
- cites
file:linethat you opened yourself; - names its cost: the wrong change it invites, or the wrong conclusion a reader draws. "This could be cleaner" names neither, and goes;
- holds where the shape stops: it does not ask for a shape that nothing in this project would enforce, or for types around a one-file script;
- respects a decision: what the project decided on purpose is a finding only where the code shows the cost. Then name the decision it reopens.
Then have each one attacked, by the most independent reader at hand:
- A model from another vendor, where the user or the project already works with one: a second CLI, a review tool. A model from the same vendor shares your blind spots.
- Otherwise a subagent with a fresh context.
- Otherwise you, arguing the other side.
Give that reader the code and the bare claim, not your argument, and ask it to refute the claim. It reads and writes nothing, like every helper in step 2: a reader that cannot be held to that is not used, and you argue the other side yourself. Run a second CLI read-only and with its input closed: one that waits for a prompt never answers. A refuted finding goes. One you still hold against an objection is reported with the objection. One round: a round of critique that only rephrases is the last round.
6. Report
In the chat. Write a file only when the user asks for one.
- The answer to the focus, in a sentence or two.
The findings, ranked by what they cost, and among equals by how far the mistake travels: out of the repository (a public API, stored data, another system), then across modules, then inside one. A race that loses an order inside one module comes before a stale sentence in a public README. For each:
- the stream and what is wrong with it, with
file:line; - what it costs;
- what would hold it (a join, a shape, the spring, the commit, a deletion) and what enforces that. Where the enforcer is a compiler setting or a lint the project does not run yet, say that the prepare skill sets it up.
Give the few worth acting on first in full, and the rest one line each. A report nobody reads to the end changes nothing. Where the topic is clean, say so, and name what you checked to know it.
- the stream and what is wrong with it, with
- Rows for reshape: the findings that are checks, in the plan table of the reshape skill, with its columns (check, kind, becomes, held by, tests), so the user can hand them over as they are. A finding that is not a check gets no row.
- What the survey did not read, so no silence is taken for a clean result.
The survey ends there. The user picks what happens next: reshape for the rows, prepare for the enforcers, a task of its own for the rest.
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.

