verify-assumptions-first
Skill1 starsChanged 38 days agoMITView on GitHub
Use when starting a large effort, such as a build of several files, a pipeline, a migration, an integration with an API, a tool or a data file not yet looked at, or any plan that rests on what a file, a service or the brief is said to be. Not for a small change, a one-line fix, or a question.
In this repository4 other files it gives its agents.
- .claude/CLAUDE.mdCLAUDE.md
- finish-taskSkill
- learnSkill
- reuseSkill
The file
We split this skill into the rules it gives an agent
6 rules. About 0 are common, 5 are distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈5
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
- Distinctive: line 12
Do the steps in this order, each as an act of its own that the user can see.
- Distinctive: line 25
Write a message to the user that lists, before any of it is checked, everything the plan takes for granted: what the inputs are (their …
- Distinctive: line 47
Run it. If it does not work, the approach is wrong, and that was cheap to learn.
- Distinctive: line 49
Do not write the rest of the build in the same call, or before this has run, however clear the whole plan is by now.
- Distinctive: line 55
Add one part in one call, then run everything that exists so far.
- Very specific: line 61
When the build is done, invoke the compound:finish-task skill (Skill tool) for the end-of-task routine.
≈ “Common” is an estimate: at least 1 in 5 of the 350,495 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
Verify the assumptions first
A large build that rests on one false assumption is rebuilt. Find the false one before anything is written, prove the approach on the smallest thing that can prove it, and only then build the rest.
Do the steps in this order, each as an act of its own that the user can see. A step that was merged into another, done in silence, or left out because the plan already looked clear was not done. The first thing after reading this is the Skill call of step 1; the second is the message of step 2. No other tool call comes before them.
1. Look for existing work
Invoke the compound:reuse skill with the Skill tool, before any other tool call. What
already exists changes the plan. A new or empty project is no reason to skip this: the
skill also looks at the user's other projects and earlier requests.
2. State the base assumptions
Write a message to the user that lists, before any of it is checked, everything the plan takes for granted: what the inputs are (their format, their fields, their size), what each API, tool or command does and that it is there, what the environment allows, and every fact the request itself states. What the user said about a file or a system is an assumption too, not a finding.
3. Check each one with a real call
The real file, the real API, the real command. Not memory, and not documentation alone. Open the file and look at its first records. Call the endpoint once. Run the tool with its version or help flag. Write one line for each assumption: it holds, it is false, or it could not be checked and why.
4. Say what was false
Tell the user at once which assumptions were false and how the plan changes. If a false one changes what was asked for, and not only how to build it, ask before building.
5. Build the smallest thing that proves the approach
One path, end to end, on the real input: the fewest lines that read the real data and produce one real result. The first call that writes a file writes this and nothing else. Run it. If it does not work, the approach is wrong, and that was cheap to learn.
Do not write the rest of the build in the same call, or before this has run, however clear the whole plan is by now. Having checked the assumptions is not having proved the approach.
6. Build out, one addition at a time
Add one part in one call, then run everything that exists so far. Go on only when nothing broke. A failure now points at the part just added. Writing every remaining file at once and testing at the end is not this step.
7. Finish
When the build is done, invoke the compound:finish-task skill (Skill tool) for the
end-of-task routine.
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.

