skills/planning-with-files/SKILL.md
Persistent markdown files as working memory for complex multi-phase tasks.
npx skillsauth add notque/claude-code-toolkit planning-with-filesInstall this skill globally with one command. Works with Claude Code, Cursor, and Windsurf.
3 of 9 scanners reported clean
Some scanners were skipped, did not run, or reported a non-clean status. Review each row below.
This skill uses persistent markdown files as external memory to execute complex, multi-phase tasks without context drift. Files serve as the single source of truth for goals, progress, and decisions. Re-read files before major decisions to ground work in written commitments rather than fallible working memory.
The workflow consists of four phases:
This skill is mandatory for tasks with 3+ phases, research requirements, or risk of context drift after many tool calls.
Goal: Establish written plan before any execution begins.
Step 1: Assess complexity
Determine if planning is needed:
Why this matters: Creating a plan takes 30 seconds but saves hours in rework. Plans prevent mid-task goal drift by anchoring decisions to written commitment. Skip this step only for single-file edits or lookups answerable in one response.
Step 2: Create task_plan.md
# Task Plan: [Brief Description]
## Goal
[One sentence describing the end state]
## Phases
- [ ] Phase 1: [First phase]
- [ ] Phase 2: [Second phase]
- [ ] Phase 3: [Third phase]
- [ ] Phase 4: [Review and deliver]
## Key Questions
1. [Question to answer before proceeding]
## Decisions Made
- [Decision]: [Rationale]
## Errors Encountered
- [Error]: [Resolution]
## Status
**Currently in Phase 1** - Creating plan
Gate: task_plan.md exists with goal, phases, and key questions defined. Proceed only when gate passes.
Goal: Collect all information needed before execution.
Step 1: Re-read plan
Open task_plan.md and read it completely. This is mandatory, not optional.
Why this matters: After ~50 tool calls, memory degrades. Re-reading restores focus and prevents drift. Claims like "I remember the goal, no need to re-read" are the primary cause of failed complex tasks.
Step 2: Gather information
Search, read, explore. Store all findings in notes.md:
# Notes: [Topic]
## Sources
### Source 1: [Name]
- Key points:
- [Finding]
- [Finding]
## Synthesized Findings
### [Category]
- [Finding with context]
Why separate files: Context window is ephemeral. Files are persistent. Writing findings to notes.md immediately ensures they survive context compression. Reference the file by section header when needed rather than keeping all content in working memory.
Step 3: Update plan
Mark Phase 2 complete. Log any decisions made. Update status line.
Gate: All key questions from Phase 1 answered. Findings stored in notes.md. Proceed only when gate passes.
Goal: Build the deliverable using gathered information.
Step 1: Re-read plan and notes
Read task_plan.md first, then notes.md. Both reads are mandatory before generating output.
Why this matters: Phase transitions are high-risk points for context drift. Two reads ground execution in both original goals and current findings, preventing divergence from the stated intent.
Step 2: Create deliverable
Build the output artifact. Reference notes for accuracy. Write to the deliverable file.
Step 3: Update plan
Mark Phase 3 complete. Log any errors encountered during execution.
Gate: Deliverable file exists and addresses the goal stated in the plan. Proceed only when gate passes.
Goal: Confirm deliverable meets the plan's stated goal.
Step 1: Re-read plan one final time
Compare deliverable against original goal and key questions. Every question should be addressed.
Step 2: Verify completeness
Check all verification criteria:
[x]Why this matters: "Done" is often an assumption, not a fact. This checklist is a defense-in-depth verification gate. Marks complete only when all criteria pass, not when work "feels done" or "should be done."
Step 3: Deliver and clean up
Present the deliverable. Remove temporary scratch files. Keep task_plan.md and deliverable as artifacts.
Gate: All verification checks pass. Deliverable is complete.
User says: "Research morning exercise benefits and write a summary"
Phase 1: Create task_plan.md with goal and 4 phases
Phase 2: Search sources, store findings in notes.md
Phase 3: Re-read notes, write morning_exercise_summary.md
Phase 4: Verify summary covers all key questions, deliver
User says: "Plan the migration from REST to GraphQL"
Phase 1: Create task_plan.md with migration phases
Phase 2: Inventory endpoints, dependencies, store in notes.md
Phase 3: Write migration_plan.md with ordered steps
Phase 4: Verify all endpoints covered, deliver plan
Cause: Too many tool calls without re-reading the plan
Solution:
task_plan.mdCause: New information invalidates original phases or decisions
Solution:
Cause: Research phase produced more content than fits in attention window
Solution:
notes.md with key takeawaysCause: Required information is missing or deliverable is off-track
Solution:
Standard File Names: Plans use task_plan.md, research notes use notes.md, deliverables use domain-specific names (e.g., migration_plan.md, research_summary.md)
Required Elements: Every plan must contain Goal (one sentence), Phases (with [ ] checkboxes), Key Questions, and Status line. Without these, the plan provides zero value.
Phase Gate Enforcement: Never advance to next phase until current phase gate passes. Gates are designed to catch problems early. "Mostly done" phases cause downstream errors. Enforce gates strictly.
documentation
Document translation: quick/normal/refined modes with chunked parallel subagents and glossary support.
development
AI image generation: Gemini and Nano Banana backends; single/series/batch workflows with prompt-to-disk.
testing
Unified voice content generation pipeline with mandatory validation and joy-check. 13-phase pipeline: LOAD, GROUND, STATS-CHECKPOINT, GENERATE, HOOK-GATE, VALIDATE, REFINE, VARIETY-GATE, JOY-CHECK, ANTI-AI, CLOSE-GATE, OUTPUT, CLEANUP. Use when writing articles, blog posts, or any content that uses a voice profile. Use for "write article", "blog post", "write in voice", "generate content", "draft article", "write about".
documentation
Critique-and-rewrite loop for voice fidelity validation.