Fulcrum 4 of 18
Between understanding and the first action
After the agent has read the message, before it edits, writes or runs anything.
What goes wrong here
The agent goes straight from its own reading to action. The person never sees what the agent thought they meant, so a misunderstanding stays invisible until it has already been built.
How it compounds if nothing catches it
Unseen misreadings are the expensive kind. By the time they surface they are files, commits and a report that all have to be unwound. The person learns to supervise every step, which is the opposite of what an agent is for.
What the harness does here
The agent opens its reply with a COHERENCE_CHECK. It covers the agent's sense of the person's intent (with a certainty percentage), what it thinks the person needs from it, and how it will communicate so it is understood the first time. It also lists any new scope it heard, quoted back with its interpretation, and whatever scope is still pending from earlier. Then it carries on working, and the person steps in only if the reading is wrong. A gate (pre-tool-coherence-gate.sh) blocks action tools until that check has been printed this turn. Reading, searching and asking the oracles stay open, because looking things up helps the agent understand. For bigger or unclear asks, align works out the nested layers of intent with the person, and reflect checks the remaining scope against them.
Jonathan, on this
Set up a harness around the coherence check alone and you'll likely see 10 to 15 percent more productivity — not from the agent doing more, but from it doing less of the wrong thing.
Jonathan Haber
The pieces that act here
- /coherence-check
Per-message alignment ritual. At the top of every reply to a person's message, print COHERENCE_CHECK and state what you think they meant (with certainty %), what they need from you, how you'll phrase your answer, every task or complaint you heard, whether to run /governer, and every open thread — then keep working. The harness's UserPromptSubmit hook injects this same text on every message and its PreToolUse gate enforces it.
- pre-tool-coherence-gate.sh
A hook: a script Claude Code runs on its own at a set moment.
- /align
Figure out what the human actually wants before doing anything. Use when starting a session, when intent is unclear, when switching topics, or when the human says "/align". This is not a technical document — it's a guide for how to listen.
- /reflect
Reflect on the ASSIGNED SCOPE that is still incomplete — what the human is trying to get, where the work stands against it, and the shortest path to close the gap. Never a reflection on the agent's own performance. Use after completing significant work, before declaring a task done, or at the start of a session to align on what you are about to do.
- /how-to-talk-like-the-founder
ALWAYS ON for every communication a human will read — every chat reply, status update, report, proposal, plan, commit message, comment, or doc, with no exempt category. Holds general, load-bearing directives for how to communicate to any human (conversational synthesis grounded in a real person's intended vs actual experience, no undefined terms, exact conditions when referencing code, broken is scope not "how it works"). Also governs writing in the first person as the specific person you're working with, which additionally requires selecting from their own verbatim material — the skill's name comes from its original build, which preserved one person's own voice; the mechanism generalizes to whoever you're working with.