Plan-mode discipline
When to skip Plan, when to triple it, the exit ritual.
What plan mode is for
Plan mode is the discipline of writing down what you’re about to do before you do it. With agents, that means: agent reads the context, proposes the steps, names the rollback path, and waits for human approval before any tool call that mutates state.
Plan mode is NOT a status update mode. NOT an exploration mode. NOT a way to ask “what should I do.” Plan mode produces a plan that is going to be executed — the seriousness comes from the commitment.
The mistake most operators make: plan mode for everything OR plan mode for nothing. Both are wrong. Plan mode is a depth dial — none / light / deep — matched to the task. The threshold logic in this lesson is what turns the dial honestly.
The plan-depth ladder
Match the depth to the task. Each rung has a different time cost and a different failure mode it prevents.
| Depth | When | Time cost | Output shape |
|---|---|---|---|
| None | Effort < 30 min, ≤ 1 file, fully reversible | 0 min | Just do the work. |
| Light | 30 min – 4 hours, 2-5 files, reversible | 5 min | ~150 words: contract row, steps, first action, rollback, out-of-scope. |
| Deep | Multi-day, multi-stakeholder, irreversible side effects, decisions upstream of steps | 30 min | ~400 words: decision required, info needed, stakeholders, steps, rollback, risk notes. |
| Triple | One-way door, organizational consequence, regulatory or legal exposure | 2-4 hours, often spread over days | Decision log + plan + risk register. Multiple specialists weigh in. CIPHER and APEX both review. |
Triple depth is rare — once or twice a quarter for most operators. SBA loan decisions, patent filings, hiring decisions, irreversible architectural commitments. The cost of getting it wrong is high enough that the planning cost is trivial.
Five failure patterns
Plan-mode-for-everything
Plan-mode-as-status-update
The plan that never exits
The plan with no teeth
The skip-the-plan-because-it's-fast
Pattern #3 (the plan that never exits) is the most insidious. Plans always have one more edge case to consider. The exit ritual is what stops the spiral.
The exit-the-plan prompt
Three checks. All three must pass to exit plan mode and execute. If any fail, revise.
Run the exit ritual on this plan. Three checks:
1. CONTRACT ROW NAMEABLE? Which Done Means row in which sprint
contract does this plan close? Cite by file path and row name.
2. FIRST CONCRETE STEP? What is the very next thing I do? Name
it precisely — file to open, command to run, person to call.
3. ROLLBACK PATH? If this plan executes wrong, how do I revert?
For code: git. For documents: version control. For
communications: prepared retraction. For irreversible
actions: there is no rollback — flag it as a one-way door.
If all three pass, exit plan mode and execute.
If any fail, revise the plan — don't execute against an
under-specified plan.Stops analysis paralysis. The plan is ready when these three things are nameable.
The exit ritual is intentionally short. Most operators over-iterate plans because each iteration reveals new detail. The three checks ask: do I know what closes this? Do I know what to do first? Do I know how to recover? Three yes answers = ship.
Three plans, three depths
A real morning across three ventures. Patent reword (none), QC dirt-path fix (light), SBA loan repoint (deep). Same operator, three depths.
The threshold logic in action. Inline / light / deep — matched to task complexity.
- operator
I have three things this morning: (1) reword three sentences in the QC patent draft, (2) implement the dirt-path P1 fix on QC, (3) figure out how to repoint the SBA 7(a) loan from Hickory to Denver NC for MHG. Which need plan mode?
- agent
Walking each through the threshold:
(1) Reword three sentences in the patent draft. Effort < 15 min. Files: 1. Reversible: yes (git). → No plan mode. Just do the edit.
(2) QC dirt-path P1 fix. Effort: ~2 hours (TDD rhythm: failing test, implementation, boundary cases, full suite). Files: 3-4 (
tests/test_dirt_path.py,src/qc/fsm/perception/pair_window.py, possibly events module). Reversible: yes (git). Sprint contract: yes (Phase 0 retry contract Done Means row). → Plan mode, light. 5-minute plan: name the test, name the API, name the boundary cases, exit.(3) SBA 7(a) loan repoint Hickory → Denver. Effort: multi-week. Files: not code — application docs, lender comms, grant assignments, decision log. Irreversible: SBA timelines and lender relationship are slow to recover. Cross-domain: legal (CIPHER), MHG (SUMMIT), commercial RE (Sarah Cooley). → Plan mode, deep. This is the one that earns 30 minutes of planning. Multiple stakeholders, irreversible side effects, structural decisions about whether to keep the loan or open a new one.
- operator
Light plan for the dirt-path fix. Five minutes.
- agent
## Plan: QC dirt-path P1 fix **Contract:** QC Phase 0 Gate Retry (memory/sprint-contracts/active.md) **Done Means row this satisfies:** "Dirt-path off_board outcomes fire on pair-window expiration" ### Steps 1. Write failing test in `tests/test_dirt_path.py` — assert off_board fires when pair window expires without a landing event. 2. Run test, confirm failure for the right reason (missing PairWindow.tick method). 3. Implement minimum: PairWindow.tick(t_ms) returns OffBoardEvent for any throw past timeout. 4. Run test, confirm green. 5. Add boundary cases: bag-lands-before-timeout (no off_board), multiple-throws-independent-timeouts, exact-boundary timing. 6. Run full `tests/test_pair_window.py` suite — confirm 14/14 green (10 existing + 4 new). ### First concrete step Open `tests/test_dirt_path.py` (does not exist yet) and write the first test. ### Rollback path If any test in the broader suite breaks: revert the implementation, keep the test (as known-broken), open a postmortem. ### Out of scope Front-cam integration, ArUco recalibration, v16 retraining.Ready to exit plan mode. Approve or rewrite?
- operator
Approved. Execute. Now do the deep plan for the SBA repoint. Take the 30 minutes.
- agent
## Plan: MHG SBA 7(a) loan repoint — Hickory → Denver NC **Status:** Hickory site DEAD as of 2026-04-28 (Denver pivot). SBA 7(a) was Hickory-scoped per memory/projects/mile-high-golf.md. **Stakeholders:** Mike (CEO), Sarah Cooley (Ascent RE), SBA lender (TBD — was Hickory-area), Curtis Woodie, Olivia Merlock (GM). **CIPHER concern:** Lender relationship implications, timeline reset risk. ### Decision required upfront A. Repoint the existing SBA 7(a) application to Denver NC site B. Withdraw and re-apply with a Denver-aware lender C. Pause SBA path entirely; pursue alternative financing This decision is upstream of any plan steps. Without it, planning is premature. ### Information needed for decision 1. Current SBA application status (filed? underwriting? approved-pending-site?) — Sarah 2. Lender flexibility on site change (some accept, some require restart) 3. Denver site lender preference (Lake Norman corridor lenders may differ from Hickory) 4. Timeline cost of B vs A (re-application typically 60-90 days) 5. Grant entanglement (Lincolnton TIP + Facade — likely lost; doesn't gate this decision but worth surfacing) ### Proposed plan steps 1. Sarah call this week — get current SBA status + lender's stance on site change. 2. CIPHER review — lender contract terms, what changes when site changes. 3. Decision log at `07-Decisions/2026-MM-DD-sba-repoint-or-restart.md` — three-option SWOT, decision recorded. 4. Action: execute A or B or C. Each has a separate sub-plan that opens once the decision is made. ### Out of scope (now) - DDI buildout rework (terminated with Hickory) - Lincolnton TIP/Facade (separate decision, likely dead) - Denver landlord LOI (separate sprint contract) ### Rollback path This is mostly information-gathering until step 4. Steps 1-3 are reversible. Step 4 is the irreversible commit — postmortem if executed wrong. ### Risk notes - SBA timelines are long; the 60-90 day cost of restart could push MHG launch. - Lender relationship in Hickory area may not transfer to Lake Norman lenders. - If Denver landlord meeting fails, this whole plan changes. ### Exit ritual - Done Means row in MHG sprint contract: "SBA 7(a) status known, repoint decision logged." - First concrete step: book Sarah call this week. - Rollback: information-gathering only through step 3.Ready to exit plan mode. This is the kind of plan that earns its 30 minutes — irreversible side effects, multi-stakeholder, decision required upstream of execution.
Prompt kit
Three prompts: threshold check, exit ritual, mid-execution divergence. Save in your CLAUDE.md.
Task: <describe>
Walk the threshold:
- Effort estimate (minutes / hours / days)
- Files affected (1 / a few / many)
- Reversibility (full / partial / one-way door)
- Stakeholders beyond me (none / one / several)
Recommend:
- No plan mode (under threshold)
- Light plan (5 minutes — name steps, first action, rollback)
- Deep plan (30 minutes — full structure with decisions, risks,
stakeholders)
Output the recommendation with the threshold reasoning.Run the exit ritual on this plan. Three checks:
1. CONTRACT ROW NAMEABLE? Which Done Means row in which sprint
contract does this plan close? Cite by file path and row name.
2. FIRST CONCRETE STEP? What is the very next thing I do? Name
it precisely — file to open, command to run, person to call.
3. ROLLBACK PATH? If this plan executes wrong, how do I revert?
For code: git. For documents: version control. For
communications: prepared retraction. For irreversible
actions: there is no rollback — flag it as a one-way door.
If all three pass, exit plan mode and execute.
If any fail, revise the plan — don't execute against an
under-specified plan.I'm partway through executing the plan at <path>. Walk the
plan vs what I've actually done.
For each plan step:
- Done as planned
- Done with a deviation (describe)
- Skipped (with reason)
- Not yet done
If deviation > 30%, the plan is failing and a re-plan is needed.
If skipped > 0, those become postmortem-eligible if they were
load-bearing.
Output: continue / re-plan / amendment needed.Apply this
30-minute exercise. Calibrate your plan threshold. Write one deep plan. Run the exit ritual.
Calibrate plan-mode discipline
Each step takes 5–10 minutes. Progress saves automatically.
- 01List 5 tasks you did this week. Walk the threshold for each — should you have planned, light-planned, or skipped?Honesty matters here. Most operators over-plan trivial work and under-plan irreversible work. Calibrate against your own pattern.
- 02Pick one upcoming task that earns deep planning. Write the deep plan.Use the structure: decision-required, information-needed, stakeholders, steps, rollback, risk notes. ~400 words.
- 03Run the exit ritual on the deep plan. Confirm all three checks pass before executing.If any check fails, revise. Don't execute against under-specified plans — that's the discipline.
- 04Save approved plans to docs/plans/active/<contract-slug>.md and reference from your sprint contract.Plans that don't bind execution are writing exercises. The contract reference is what makes them real.
- 05Mid-week, run the divergence check on any active plan. Catch drift before it becomes a postmortem.Drift > 30% means the plan was wrong, not that you executed badly. Re-plan, don't push through.