observe-specify-build-verify
Skill0 starsChanged 5 months agoMITView on GitHub
Run a fail-closed Observe-to-Verify artifact gate.
In this repository84 other files it gives its agents.
- AGENTS.mdAGENTS.md
- docs/AGENTS.mdAGENTS.md
- shep/AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- shep/CLAUDE.mdCLAUDE.md
- advisory-version-checkSkill
- aeo-llm-content-plannerSkill
- agency-command-centerSkill
- agency-morning-briefSkill
- agency-scorecardSkill
- agent-runbookSkill
- agent-usageSkill
- autonomy-budgetSkill
- backlink-pr-angle-generatorSkill
- brand-kit-from-urlSkill
- browser-automationSkill
- building-autonomous-agentsSkill
- capability-mapSkill
- case-study-engineSkill
- clickup-apiSkill
- client-dashboardSkill
- client-health-scoreSkill
- client-onboarding-osSkill
- cli-llm-routingSkill
- codebase-mapSkill
- command-from-your-phoneSkill
- competitor-content-gap-analysisSkill
- connection-monitorSkill
- content-brief-generatorSkill
- determinism-patternSkill
- dispatch-and-briefSkill
- division-operating-modelSkill
- dreamingSkill
- durable-stateSkill
- entity-topical-authority-mapperSkill
- first-real-jobSkill
- founding-circle-ai-securitySkill
- gbp-competitor-auditSkill
- goal-modeSkill
- host-your-agentSkill
- html-reportsSkill
- instinct-bridgeSkill
- evolveSkill
- learn-evalSkill
- learnSkill
- promoteSkill
- skill-createSkill
- internal-linking-auditorSkill
- keyword-deep-diveSkill
- ladder-auditSkill
- lanes-momentum-frameworkSkill
- launch-websiteSkill
- llm-citation-auditSkill
- migration-seo-paritySkill
- migration-visual-checkSkill
- model-gear-routerSkill
- multi-account-gatewaySkill
- multi-model-councilSkill
- night-shiftSkill
- novaSkill
- The first 60 shown.
The file
We split this skill into the rules it gives an agent
8 rules. About 0 are common, 7 are distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈7
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 20
Use this small workflow to turn an owned reference and an approved request into a reviewable artifact.
- Distinctive: line 24
Use for a bounded exercise where a candidate must be checked against an approved specification.
- Distinctive: line 25
Use when missing artifacts, missing evidence, spec drift, or altered files must stop the flow.
- Distinctive: line 26
Do not use it as a substitute for legal permission, browser QA, security review, or a production release gate.
- Distinctive: line 33
No installs or network access are needed.
- Distinctive: line 37
Observe: Read only the permitted local reference.
- Very specific: line 41
Gate: Run the deterministic programmatic gate separately from any fallible LLM review.
- Distinctive: line 70
If browser behavior matters, perform a separate local browser check against the synthetic page and record that evidence.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 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
Observe, Specify, Build, Verify
Use this small workflow to turn an owned reference and an approved request into a reviewable artifact. The included gate is deterministic and fail-closed. It is a static contract check, not proof of browser behavior, autonomous agent quality, or permission to publish.
When to Use
- Use for a bounded exercise where a candidate must be checked against an approved specification.
- Use when missing artifacts, missing evidence, spec drift, or altered files must stop the flow.
- Do not use it as a substitute for legal permission, browser QA, security review, or a production release gate.
Prerequisites
- An original request and an
APPROVEDspec with required criteria. - Reference material that is synthetic or owned, with permission to use it. Do not copy assets without permission or bypass authentication or DRM.
- A private
work/directory for receipts and personal output. Never put personal information in tracked fixtures. - No installs or network access are needed. The example uses only Python standard library code.
Procedure
- Observe. Read only the permitted local reference. Record what is relevant and what is out of scope. Completion: the source and permission boundary are written down.
- Specify. Convert the original request into an approved spec with stable criterion IDs and exact expected evidence. Completion: status is
APPROVEDand every required criterion is testable. - Build. Give the builder the approved spec and permitted inputs, not the original request. Save the candidate under private
work/until review. Completion: the candidate and its SHA256 are available. - Verify independently. Give the verifier the original request, spec, candidate, and evidence. The verifier checks the artifact and can catch spec drift. Completion: evidence binds to both hashes and covers every required criterion.
- Gate. Run the deterministic programmatic gate separately from any fallible LLM review. Completion: receipt says
PASS,REQUEST_CHANGES, orBLOCKED; missing or invalid proof is alwaysBLOCKED. - Release only after approval. A passing fixture does not grant live authority or publish anything. Completion: a human or separately authorized release process records the next approval.
Safe role prompts
Builder:
Build one candidate from this approved spec and permitted local inputs only. Do not infer unstated requirements. Write the candidate to private
work/. Do not publish, contact external systems, bypass access controls, or copy assets without permission. Return the artifact path and SHA256.
Independent verifier:
Review the original request, approved spec, candidate, and evidence independently. Check every required criterion and report omissions, drift, invalid evidence, and hash mismatches. Return only PASS or REQUEST_CHANGES for a complete review. If the artifact or evidence is absent or invalid, return BLOCKED. Do not publish or treat a role prompt as a sandbox or legal guarantee.
The information boundary is a review design, not a security boundary. Clean-room builder inputs and independent QA are distinct controls, but neither role prompt prevents leakage or authorizes use of protected material.
Quick Reference
From examples/observe-specify-build-verify/:
- macOS/Linux tests:
python3 -m unittest discover -s tests -v - Windows tests:
py -3 -m unittest discover -s tests -v - macOS/Linux demo:
python3 demo.py - Windows demo:
py -3 demo.py
Expected demo verdicts are incomplete: REQUEST_CHANGES, corrected: PASS, and missing-evidence: BLOCKED.
Pitfalls
- A marker or parsed anchor proves only the static content contract. It does not prove rendered visibility, CSS behavior, navigation, accessibility, or browser behavior.
- If browser behavior matters, perform a separate local browser check against the synthetic page and record that evidence. Do not upgrade a static PASS into a browser PASS.
- Craft Anything is inspiration only. Its actual material describes Study, Build, Check, and Ready; it does not prove that two teams never saw the original, is early alpha rather than a benchmark, and is an app/CLI/MCP repository rather than this ready-made skill.
- A fallible LLM reviewer may suggest changes, but the deterministic gate remains the authority for hash, schema, required-evidence, and fixture-contract checks.
Acceptance and Rollback
Accept only when tests pass, the corrected candidate returns PASS, receipts bind the approved spec and candidate hashes, and any required browser or human review is separately recorded. To roll back, delete private work/ receipts and discard the candidate. Do not publish or mutate the source reference.
Verification
Run the example tests and demo with the commands above. The tests cover absent artifact, absent evidence, altered candidate hash, missing criterion evidence, invalid spec, spec drift, an intentionally incomplete candidate, and a corrected candidate. Review the JSON receipts under private work/; they are exercise evidence only.
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.

