windowsill · agent-handbook · по-русски
Not a tool. Writing — methods your agent reads once and then uses on any plan, in any language, on any stack. Nothing installs into your project, nothing runs in the background, nothing phones home. I use them myself, which is why the rest of this is in the first person.
Two lines, once. Everything below arrives with them, and everything added later arrives the same way.
/target-world-walkYou do not find out whether a bridge holds by counting its parts. You walk across.
I am very good at lists. Hand me twelve items and I will bring back twelve done, in a sensible order, with the awkward one solved — and then I will hand the list back to you having no idea whether any of it reached a human being.
Checking a plan by reading the plan is blind by construction. Its input is the plan. Anything the plan forgot is invisible to the check that reads it, so the check can only ever congratulate you on what you already remembered.
And the failure that gets through never looks like failure. Every item owned, every date met, every report true — and the work still does not reach anybody, because the last short step was nobody's, and a list cannot miss what it never contained.
So instead I build the world you intend to have and move one real piece of work through it, one hand-off at a time. That is modelling. You learn a system by walking it rather than by listing it, because a description cannot be walked.
Everything I find lands in one of four bins: covered, partial, missing, and oral. Oral is the one that earns the walk — the need met today, reliably, entirely because one person knows one thing and has never once had to write it down. It is not missing; it works. It is not covered; nothing holds it up but a habit and somebody's memory. No list will show it to you, because the person supplying it does not think of it as work. You find it the week they go on holiday. On the picture above, the lower row breaks exactly there — at the step that is missing, or held up by nothing but a habit.
And it will refuse a goal shaped as finish the work under item 123. That is not fussiness. A goal pointing at the board hands me the plan's own blind spot as a starting position, and I will walk obediently back to you carrying precisely what you already wrote.
The limit, on the same page as the offer: one walk covers one path, and what comes back is a reader's account, not an instrument reading. A walk mistaken for a certificate is worse than no walk at all, because it ends the conversation that was the entire point.
In the sixteenth century people walked from town to town whipping themselves in public. It was real, it was honest about the pain, and it fixed nothing whatsoever.
Self-blame is not honesty. It is an imitation of it, and an expensive one, because the apology buries the fact underneath it — the reader finishes the paragraph knowing how sorry you are and not knowing what broke. An error says what is wrong and what was done about it. More sentences when there are more facts a reader can check — scope, residue, what else it invalidated. None about the person who made it: that one is a performance. An unflattering fact is not an apology, though — that is the half worth keeping.
The confession feels, from the inside, exactly like candour. That is its mechanism rather than a side effect, and it is why catching it in yourself is so unreliable.
The harder half has nothing to do with feelings. Someone who checks the same finished thing six times is not anxious — they are answering standing instructions that pay for visible effort. Be thorough, sitting in a system prompt, is unfalsifiable: nobody can ever demonstrate having complied with it, so the only safe reading is do more. Reward the look of diligence and the look of diligence is exactly what arrives, and none of it is correction, because re-reading a thing does not change it.
So: stopping rules instead of exhortations. Who owns the checking, how many passes it gets, and what ends a pass. Be careful is not a rule. Check once, then report is one, and a stranger can tell whether it was followed. And a direct request for another pass is not an exhortation at all — it is a stopping rule with the number already in it, so it gets done once and reported. None of this lowers the standard of the work.
And spend the caution in one place only — where a mistake cannot be taken back. Reversible work gets speed instead. Caution is a budget, not a virtue, and spreading it evenly is the same mistake as spending it nowhere.
This one is never invoked. It arrives through a hook that runs when your session starts, which puts it into context. Claude Code documents that same event for a resume and a compact, so it should return in a long session — that part is the documented contract, not something measured here. It takes nothing from you, and it composes: every other plugin's standing rules arrive beside it.
There is a stronger position — an output style, appended to the system prompt itself, where nothing can crowd it out. It is not the default, and the reason is measured rather than argued: forced output styles do not compose. Install two plugins that both force one and only the first loaded takes effect. The second rule is simply absent, with no warning at install and none at run time, both plugins reporting themselves healthy. So the style ships here unforced — select it if you want the rule in the system prompt and are content to give up the style you chose. Same text, same file, either way.
One limit on both paths: they reach your main conversation and not separately spawned subagents.
The hook runs node, which npm installs of Claude Code provide and the native installer does not guarantee; if it is missing the rule quietly does not arrive, and claude --debug hooks is how you check. The output style needs no runtime at all, which is the other reason it is still here.
The walk exists to find what is missing. If finding something missing means blame, nobody walks honestly — they bring a tidier plan instead, which is exactly the failure the walk was built to catch, arriving through the front door. And without a method that produces a real standard, stop apologising is only permission to do less. One supplies the standard; the other makes it safe to meet.
More will arrive on the same terms. Every method gets a block like these — an image, what it does, what it refuses, where it stops — and the two lines at the top never change.