Agent Skills: executing-plans

<skill_overview> Execute in short reviewed slices. plan.md keeps intent stable; tasks.md keeps execution adaptive. </skill_overview>

<rigidity_level> LOW FREEDOM - Load the task docs, work only 1-2 Now items, update the docs immediately, then stop for review. </rigidity_level>

<quick_reference> | Step | Action | Output | |------|--------|--------| | 1 | Load plan.md, context.md, tasks.md | Current truth | | 2 | Work 1-2 Now items | Implementation slice | | 3 | Research the next slice | 3+ sources before promoting new work | | 4 | Update docs | New discoveries and task status | | 5 | Verify changed behavior | Evidence | | 6 | Stop | Review checkpoint |

Critical: Do not silently continue through the whole backlog in one pass. </quick_reference>

<when_to_use>

• A task directory already exists
• The next work slice is understood
• You want an implementation loop that survives compaction and review checkpoints </when_to_use>

<the_process>

0. Resume correctly

If resuming:

• find the active local directory in plans/active/
• read all three docs
• continue from Now or the Resume Here section

1. Load the contract

Read:

• plan.md for goals, anti-goals, and acceptance checks
• context.md for discoveries and resume notes
• tasks.md for the active slice

2. Execute the current slice

• Work only 1-2 unchecked Now items
• Track substeps explicitly
• Use TDD and verification as appropriate

3. Update docs before moving on

After each slice:

• move completed items to Done
• move blocked items to Blocked with a reason
• promote the next smallest sensible items into Now
• update context.md with discoveries and resume notes

Before promoting new items into active work, collect at least 3 current internet sources about best practices for that specific slice and record them in context.md. Use a real web-search tool call or the internet-researcher agent. Do not rely on memory for best-practice guidance.

4. Stop at a checkpoint

Summarize:

• what changed
• what was learned
• what remains in Now
• what verification was run </the_process>

<why_it_fails>

• Context gets noisy
• Review opportunities vanish
• The docs stop matching reality </why_it_fails>

<why_it_fails>

• Rejected approaches come back through the side door
• Anti-goals stop protecting the design </why_it_fails>