est7 Deep Plan
Structured deep analysis with script-enforced output gates. Unlike prompt-only self-checks, this skill validates thinking quality programmatically — the validator script parses your output and rejects incomplete analysis.
When to use this vs plan-code-workflow
- plan-code-workflow: organizing what to build and how — Plan mode → Code mode
- deep-plan: rigorous reasoning before deciding what to build — when the decision itself is the hard part
Use deep-plan first, then hand off to plan-code-workflow for execution.
Phase 1 — Complexity Assessment
Classify the task:
| Level | Criteria | Min Steps |
|---|---|---|
| simple | Single file/function, no cross-module impact | 3 |
| medium | Multi-file/module, design choices needed | 5 |
| complex | Cross-service, architecture tradeoffs, migration | 8 |
Output the complexity level and your initial step count estimate. This determines the validator's minimum step threshold.
Phase 2 — Sequential Thinking
Output numbered thinking steps. Each step MUST use this exact format:
### Step N/M [tag]
- **Focus**: What this step analyzes (one concern per step)
- **Findings**: Observations, evidence, conclusions
- **Confidence**: high | medium | low
- **Next**: What the next step addresses (or "→ Gate" on final step)
Tags
| Tag | When to use |
|---|---|
analysis |
Examining a specific aspect |
decomposition |
Breaking problem into sub-problems |
evaluation |
Assessing an approach or alternative |
comparison |
Comparing options side-by-side |
revision |
Revising a previous step — use [revision of Step X] |
branch |
Exploring an alternative path — use [branch: id] |
synthesis |
Combining findings into a conclusion |
Rules
- One concern per step. Do not combine multiple analyses.
- Adjust M freely. Add steps as understanding deepens; reduce if a path is clearly simpler.
- Revise explicitly. When new evidence contradicts an earlier step, tag as
revisionand cite which step. - Address low confidence. Any step with
Confidence: lowmust be followed up by a later step that resolves the uncertainty. - Terminate with "→ Gate". The final step's Next field must say "→ Gate" to signal readiness for validation.
Required coverage
Your steps must collectively cover all of these (order is flexible):
- Problem definition — core problem, constraints, success criteria
- Decomposition — sub-problems (MECE: mutually exclusive, collectively exhaustive)
- Evaluation — at least 1 alternative approach assessed (or explicit justification why only 1 viable path exists)
- Tradeoff comparison — tabular comparison on ≥3 dimensions (when alternatives exist)
- Execution steps — 4–10 ordered, concrete, verifiable actions
- Risks — issues and concrete mitigations
Phase 3 — Validation Gate (Script-Enforced)
After completing all thinking steps, synthesize into a plan document and run the validator.
Step 3a — Write the plan to a temp file
Write to /tmp/deep-plan-output.md containing both your thinking steps AND the plan:
complexity: <level>
(All your ### Step N/M [tag] blocks from Phase 2)
## Plan: {task title}
### Problem Definition
- **Core problem**: ...
- **Constraints**: ...
- **Success criteria**: ...
### Analysis Summary
Key findings referencing step numbers.
### Execution Plan
1. Step — description → verification method
2. Step — description → verification method
...
### Risks
- Risk → mitigation
### Open Questions
- [ ] Assumptions needing confirmation
Step 3b — Run the validator
python <skill-path>/scripts/validate_plan.py /tmp/deep-plan-output.md
The script checks 8 gates:
| Gate | What it checks |
|---|---|
step_format |
Every step has Focus, Findings, Confidence, Next |
step_tags |
All tags are from the valid set |
coverage |
All required topics addressed |
termination |
Final step ends with "→ Gate" |
low_confidence_addressed |
Low-confidence steps resolved later |
plan_sections |
Plan has all required sections |
execution_verification |
Every execution step has a verification method |
minimum_steps |
Step count meets complexity threshold |
Step 3c — Handle results
- All gates pass → proceed to Phase 4
- Any gate fails → read the failure details, add thinking steps to fill the gaps, rewrite the plan, rerun the validator. Repeat until all gates pass. Do NOT proceed with failing gates.
Phase 4 — Review Gate (Human)
Present the validated plan to the user. Then pause with AskUserQuestion:
- "Approve — proceed to execute"
- "Adjust — specify what to change"
- "Abort — discard this plan"
Do NOT proceed without explicit approval.
Phase 5 — Persist and Execute
Only after approval:
- Write plan to
plan/YYYY-MM-DD_HH-mm-ss-<slug>.mdwith frontmatter:
---
mode: plan
task: <task title>
complexity: <simple|medium|complex>
created_at: <ISO8601>
validated: true
---
- Execute the plan step by step.
Session Continuity
- "adjust", "modify", "refine" → update plan, re-validate, re-enter Phase 4
- "new", "different task" → restart from Phase 1
- Uncertain → ask the user