A skill for the moment you hand a job to an agent and walk away.
Two things go wrong when you do that. The agent gets timid — it does 20% of the work and comes back with a menu of options you have to read. Or it gets confident — it reports success it never verified, and you find out three days later that "installed" never meant "running."
/direct is one page of instructions that addresses both, plus a small mechanism that
makes it get better at your codebase every time it's wrong.
Act. No "shall I?" for reversible steps that obviously follow from the request. No menu of options it could have simply tried. It crosses the boundary of the task instead of stopping to describe what's on the other side.
Prove. A claim without a command output behind it isn't a claim. Not "the config is set" — probe the socket. Not "the tests should pass" — run them. It's told to prefer a check that could have failed, because a check that can't fail proves nothing.
Stop for exactly four things. Everything else, it proceeds. The four are: irreversible or outward-facing acts in your name; grading its own homework (weakening the test that would have caught it); anything physical; and destroying work it didn't create. Naming them narrowly is what makes "proceed" safe to mean literally.
Learn. After each run it appends what it got wrong to references/lessons.md, and
reads that file before planning the next one.
Most "learning" agent setups fail the same way: an open invitation to journal produces a file full of "the task went well," nobody reads it, it grows without bound, and it dies. This one is deliberately narrow.
- It only writes on failure. The harvest condition is a closed list: a verification you expected to pass failed; a doc or your own prior claim turned out false; a check went green that wasn't real; the owner corrected you; a guard denied you; or something cost more than ~20 minutes that one sentence would have prevented. A clean run writes nothing. That's what keeps the file worth reading.
- Fixed entry shape. Believed → Reality → Rule → Cost → Fires. The Rule is one imperative sentence a future session can apply with none of today's context. The rest is there so a human can audit whether the rule is justified.
- Repeats increment, they don't duplicate. Same trap twice bumps
Fires: 2. - Promotion at 3 fires (or one hour-plus incident): the rule graduates out of the log and into the skill body or the verification playbook, where it's read unconditionally.
- Bounded. Capped at 40 live entries; pruning removes only entries that never re-fired and whose subject no longer exists on disk. Bounded memory that gets read beats unbounded memory that gets skipped.
- It can't edit its own past. Never rewrite a lesson to make a run look better, and the no-grading-your-own-homework rule explicitly binds the harvest condition itself.
The result is a skill whose value curve is the opposite of a prompt's: the parts that are specific to your machine get written by your machine's failures, not guessed at by whoever wrote the file.
Copy the direct folder into any skills directory Claude Code reads:
# every project on this machine
cp -r skills/direct ~/.claude/skills/
# or just this project
cp -r skills/direct .claude/skills/Windows PowerShell:
robocopy skills\direct "$env:USERPROFILE\.claude\skills\direct" /EThen start a session and say /direct — or "do wonderful things", "finish it", "do it
yourself". The description triggers on all of them.
references/verification-playbook.md ships as a template, and it's the half of this
skill that can't be shared. "Prove it" means something different in every codebase. Fill in:
- the health check for each service you run, phrased as the proof, not the thing;
- the exact test / build / typecheck / e2e commands that constitute evidence here;
- the signals you know lie in your environment.
A /direct with a filled-in playbook is dramatically more useful than one without. Leave
lessons.md empty — your first real entry arrives on its own within a few tasks.
Things this file does on purpose, which you may want to steal:
- The stop-list is short and enumerated. Vague caution ("be careful with risky actions") makes an agent hesitant everywhere and safe nowhere. Four named categories make "everything else, proceed" a sentence you can actually mean.
- The self-check is adversarial. "Verify with something that could have failed" is the single highest-leverage line in the file.
- Guard denials are findings, not obstacles. The skill forbids routing around a PreToolUse guard, and tells the agent a denial is often the most interesting result of the run — which converts a frustration into a reported signal.
- Learning is gated on falsification, not on activity. This is the whole trick.
MIT. See LICENSE.