🔄 Agentic Methodology · Execution Discipline

Verify before claiming done — evidence over claims

Fresh-command proof for every completion claim; 'it should work' is a confession.

foundation~20 minAgent Power UsersAI-Native EngineersTeams Shipping With Agents

Steps

1Run the actual verification command fresh at claim time, not from memory
2Paste real output into the completion report — truncated honestly if long
3Re-test edge cases named in the original task, not just the happy path
4Distinguish 'compiles', 'tests pass', and 'works' as three separate proofs
5If anything remains unverified, say exactly what and why — never gloss
6Treat any 'should' in your own status report as an unfinished task marker

Common Pitfalls

▲Agents declaring success on stale terminal output
▲Confident summaries laundering unverified assumptions

Commands

Install with skills CLI

$ npx skills add aniruddhaadak80/skills --skill execution-discipline-verification-before-completion

Install globally

$ npx skills add aniruddhaadak80/skills --skill execution-discipline-verification-before-completion -g

Tags

#verification#honesty#workflow#agentic-methodology#execution-discipline

Install this skill

$ npx skills add aniruddhaadak80/skills --skill execution-discipline-verification-before-completion -a opencode

SKILL.md

# Verify before claiming done — evidence over claims

> Fresh-command proof for every completion claim; 'it should work' is a confession.

## When to Use This Skill

Fresh-command proof for every completion claim; 'it should work' is a confession.

## Steps

1. Run the actual verification command fresh at claim time, not from memory
2. Paste real output into the completion report — truncated honestly if long
3. Re-test edge cases named in the original task, not just the happy path
4. Distinguish 'compiles', 'tests pass', and 'works' as three separate proofs
5. If anything remains unverified, say exactly what and why — never gloss
6. Treat any 'should' in your own status report as an unfinished task marker

## Common Pitfalls

- Agents declaring success on stale terminal output
- Confident summaries laundering unverified assumptions

View source on GitHub ↗

Failing test first, watch it fail, minimal code, watch it pass, commit — no code born before tests.

🔄 Agentic Methodology·~30m

Open →

Dispatch subagents with fresh-context disciplineadvanced

Fresh implementer per task, brief files as single source of truth, two-stage review, capped fix loops.

🔄 Agentic Methodology·~40m