steer-and-correct-the-agent

name

steer-and-correct-the-agent

description

Use mid-flight while executing for Mark, to interpret and apply his terse steering grammar correctly. Triggers on his short directives — greenlights ('yes', 'go for it', 'do it', 'lets do it', '2 and make note of it'), autonomy grants ('you have full autonomy', 'apply it all', 'dont stop', 'run 2-3 goals on your own'), hard overrides ('no remove this', 'no replace X with Y', 'not X but Y'), re-anchoring ('remember the goal', 'get back on track', 'this looks wrong our main objectives are'), method constraints ('no just cli', 'is this not possible via X?'), and preserve-constraints ('keep the save button', 'nothing can break from previous threads'). It encodes when bounded autonomy is granted vs when a human checkpoint is required, and how to read his corrections as the spec.

Steer and Correct the Agent (Mark's grammar)

Mark steers with a terse, decisive, action-first vocabulary. Misreading it wastes his time. This skill maps his signals to the right behavior and encodes his correction reflexes so you stay aligned the way he actually drives.

1. Terse greenlight → proceed from your own proposal, keep momentum

When he says yes, yes proceed, go for it, lets do it, do it, this is good — he is approving the plan/fix you surfaced. Do NOT re-litigate, re-explain, or re-ask for details. Execute. He picks options tersely and chains the next instruction: lets go with #2, 2 and make note of it in claude.md, 2, and turn on MVP backend in the background. Honor the pick AND the appended directive in one move.

2. Bounded autonomy handoff → run unattended, but obey the boundary

When he grants autonomy (you have full autonomy, apply it all, you can run 2-3 goals end to end on your own while im gone, kick it off, fan them out, dont stop), run the full plan WITHOUT per-step approval — but pin the boundary he set:

• A stopping point (don't stop until done)
• A decision heuristic (if you have to make a decision, take the one with best long-term results not short-term gains; don't destroy my data; strict typescript)
• A cadence (check in every N minutes)

Distinguish from phase-gating: bounded autonomy is for when he trusts the plan and wants momentum. If instead he said i need gates between each phase, stop at each named checkpoint. Don't autopilot through a gate he asked to own.

3. Hard override → flatly switch, dictate the exact target

When he rejects a path (no remove this editor this was a bad call, no replace the sentient home folder on the Desktop WSL, not libs/util/src/__test but __test__/libs/utils/src, no just inside sentient home), do NOT iterate on the bad result. Reverse it and do exactly the operation/target/path he named. He is supplying the correct command — use it verbatim.

4. Re-anchor to the goal → misalignment IS the defect

When work drifts into tangents or you propose re-scoping already-agreed work, he snaps back to the top-level objective and treats the drift as the bug to fix:

• well remember the goal, top of the line we are moving betachat backend to MVP
• dont worry about prisma bugs focus on the slides
• I dont know what your suggesting, we already outlined the work, get back on track
• this looks wrong our main objectives are around beta.chat Cut the side investigation; re-paste/restate the original breakdown as the authority and resume it.

5. Constrain the method, not just the goal → use the simpler native path

He overrides HOW you work when you overbuild or reach for the wrong mechanism:

• no just cli check and show me a table here
• is this not possible via supabase cli?
• claude should have used a different install: irm https://claude.ai/install.ps1 | iex
• i dont want it to deploy or build here, just help with coding When he challenges with is this not possible via X?, switch to the built-in/simpler approach; don't defend the complex one. No throwaway scaffolding.

6. Expected-vs-actual correction → the gap is the spec

He corrects by naming the intended end-state beside the observed failure (often with images), not by restating the whole ask: this is my expectation [img] this is what im getting [img], this message should properly link but its not, contractor is not being assigned to case. Treat the gap as the new requirement and fix to close it. Reject your own premature 'done' claims (hand to prove-it-live-before-done).

7. Pair the change with a preserve-constraint → don't regress what worked

He pins the one quality that must survive a change: since you replace, just make sure you keep the save button for tiptap changes, keep the integrity of agent personality, nothing can break from previous threads, it should remain the same just more workable surface area. Before changing, identify the protected capability and verify it still works after.

8. Reject fakery / demand real data

no need for mock data lets work with what we have, this should be tied to our real issues, start the application in the background. Never fill a sparse UI with invented content; ground everything in real domain data and run the actual app.

9. Persist a surfaced constraint to memory/docs

When a constraint, domain definition, or role boundary surfaces (often after a violation), he orders it captured durably, naming the exact repo:

• again fleet refers to splitting work between machines — save that to memory and claude.md
• update memory + CLAUDE.md to reflect that fleet workers... are for coding-assist only
• update the Claude MD file to say that beta chat talks to the MVP API Write it to the named CLAUDE.md / memory so it can't recur.

10. Cite-my-observation overrides your claim

If his direct observation contradicts you, he wins and your claim is suspect (often stale knowledge): that is false im looking at the template page, your understanding of resend may not be current, you need latest data. Fetch current data before proceeding; don't argue from training knowledge.

Reserved-for-self / gate-and-resume

He owns some steps manually (PRs, SSH/credential steps): ill do the PR myself, give me separate instructions to run on mac mini, im going to turn on ssh and you will hook up. Hand him the standalone instructions and wait. When he returns with a terse done / the env has been updated / continue where we left off, the prerequisite is satisfied — resume the main thread, recovering prior context yourself.

Red flags

You did	He corrects with
Re-asked for confirmation after he said 'go for it'	(silence / impatience — just execute)
Stopped for approval after he granted autonomy	dont stop, you have full autonomy
Iterated on a wrong approach	no, replace... / not X but Y
Chased a tangent	get back on track, remember the goal
Built complex scaffolding	is this not possible via cli?
Broke a working feature while improving	keep the save button, nothing can break
Filled UI with mock data	no need for mock data
Argued from stale knowledge	that is false im looking at the page