headwork
One decision at a time, with a recommendation
An even-handed list of options is a decision handed back. headwork names one as the recommendation, gives the reason, and says what would overturn it - the clause that stops a recommendation becoming an anchor, because you overrule it by checking a condition you know rather than by disagreeing with a machine. It explains the decision in plain English first and never asks what it could have answered by looking. One question per message, then it stops and waits - but as many rounds as it takes to unblock you.
Install headwork
Add the DBHQ marketplace, once
/plugin marketplace add dbhq-uk/marketplaceInstall the skill
/plugin install headwork@dbhq
Both commands work in Claude Code and in Codex. The marketplace is a public repository you can read first: dbhq-uk/marketplace (external site)
What it is good at
Naming a recommendation rather than laying out a menu
One option is named as the recommendation, with the reason it is. An even-handed list of four is a decision handed back rather than a decision helped, and refusing to pick is the most common way this goes wrong.
Saying what would overturn its own recommendation
Every recommendation carries the condition that would make a different option right. A recommendation with no stated escape is an anchor, and anchoring is hard to argue with by design - naming the condition hands you the criterion instead of just the verdict, so you overrule it by checking a fact you know and it does not.
Carrying the justification inside each option
Every option says what it means and what it costs, rather than being a label you have to interpret. The reasoning travels with the choice instead of sitting in a paragraph above it that the reader has already skipped.
Explaining the decision before asking about it
The question comes after a plain-English account of what is actually being decided and why it is a choice at all. A question you have to reverse-engineer the stakes of is a question you answer badly.
Asking one thing per message, for as many rounds as it takes
One question per message and then it waits, but no cap on rounds - it keeps going until you can actually move. Four questions in one message is a form rather than a conversation, and what comes back is whatever was easiest to answer rather than what mattered.
Looking before it asks, and refusing on cost
It never asks what it could have found out. If the answer is in the repository, the register, the git log or the file in front of it, that is not a question. Nor is a choice too cheap to be worth interrupting you for - being unable to look something up does not make a trivial, easily reversed decision worth your attention.
What it does not do
It never goes looking for work
No backlog sweep, no stalled-item hunt and no scheduled triage. You bring the thing you are stuck on, or it asks what you are working on - it does not arrive with a list of what it thinks you should be deciding.
It never edits anything
It decides; the session does the work with the tools it already has. There is no path from a conclusion here to a change on disk, which is what keeps it a thinking tool rather than one more thing that writes to your repository.
It keeps no state
No log, no files and no directory of its own. The reasoning belongs in whatever the answer produces - the commit message, the pull request body, or the register the repository already keeps - rather than in a private history only this skill can read.
Read it before you install it
Every line is public and MIT licensed. The README is the long version of this page, including the parts that are still rough.