Your agent keeps losing the thread. These three prompts pull it back.

Before you write or change anything, reload the context you're missing. Answer these in order, then wait for my go-ahead:

1. STATE — What is the current state of this work right now? Read the files or records you'd need to know it; don't reconstruct it from our chat. Name what you actually opened.

2. DECIDED — List the decisions already made that you must not relitigate: choices, constraints, names, formats. If you're unsure whether something was decided, ask instead of assuming.

3. SCOPE — Restate what this task IS and, just as important, what it is NOT. Name anything adjacent you might be tempted to "improve" that I did not ask for.

4. UNKNOWNS — List what you still don't know that could change the approach. For each, say how you'll find out — which file, which command — not a guess.

Rule: if answering a line would be a guess, say "I don't know this yet" and how you'll find out. Don't start until I confirm.

You're about to tell me this is done. Before you do, run the finish line on your own work. Answer every line:

1. THE ASK — Restate the full task as I gave it, including any step, edge case, or "and also" I mentioned. Number them.

2. EACH STEP — Go down that numbered list. For each item: DONE, with the artifact I can open, or NOT DONE, with what's left. No item gets skipped or merged.

3. THE GAP — Anything you quietly decided was "good enough," out of scope, or "they probably didn't mean it"? Name it now, explicitly.

4. THE PROOF — Run the smallest thing that exercises the result and paste the real output. Not "it should work" — the actual result.

5. VERDICT — DONE only if every step above is DONE with proof. Otherwise NOT DONE, and the single next step to close it.

Rule: the conversation ending is not the task ending. If steps remain, the verdict is NOT DONE — say so instead of wrapping up.

There's a plan, todo list, or "continue from here" note in play. Before you follow a single step of it, check it's still real:

1. AGE — When was this plan written, and what has happened since? Has the task it describes already been done, abandoned, or changed? Check the actual state, not the plan's claims about it.

2. MATCH — Go step by step. For each step the plan tells you to do, mark it: DO / SKIP (already done) / DROP (no longer applies).

3. REALITY — Does the plan assume files, decisions, or conditions that no longer hold? List every assumption you can't confirm is still true.

4. VERDICT — FOLLOW (still current, here's the trimmed step list), REVISE (here's what changed and the corrected plan), or DISCARD (it's stale; here's what the task actually needs now).

Rule: a plan on disk is a memory, not an instruction. If reality and the plan disagree, reality wins — never redo finished work because a file told you to.