Compound Simple

compound-simple

Simplified compound skill for lightweight post-task reflection. Embed session learnings in ..HEAD using only your own commit range. If you made no commits, check git diff --stat` for uncommitted changes instead.

Answer:

• What files changed (in your own commits only)?
• What was accomplished?
• Was the task's Done-when condition fully met?

Step 2: Check for learnings (5 questions)

Answer each question. If yes, note what to update:

1. Something fail unexpectedly? (silent error, missing dependency, wrong assumption)

→ Note the failure and workaround in task completion note

1. Something succeed unexpectedly? (easier than expected, found shortcut)

→ Note the technique for potential reuse

1. Workaround worth noting? (non-obvious trick, configuration quirk)

→ Add to project STATUS.md or task note

1. Task description miss something? (prerequisite not listed, scope underestimated)

→ Note for future task writing improvement

1. Anything reduce work efficiency? (misleading requirements, confusing docs, awkward tool behavior, wrong parameters, tiny annoyances, or other friction)

→ Write it to the relevant project's ISSUES.md: what happened, when you hit it, and how you investigated it. Record even small friction if it slowed you down at all — minor annoyances are still worth capturing because a later model can remove the root cause.

Step 3: Check for implied tasks

If you completed an experiment/analysis, scan results for:

• "FAIL" or "below threshold" → task: refined experiment
• "N too small" → task: larger replication
• "unexpected", "mechanism unclear" → task: diagnosis
• Multi-phase plan → check phase tasks exist
• New bugs discovered → task: bug fix
• Documentation outdated → task: doc update

Step 4: Act on findings

For each compound opportunity:

| Type | Action | |------|--------| | Small update (1-3 lines) | Apply directly to file (STATUS.md, task note, ISSUES.md) | | Larger change | Create task in TASKS.md with provenance | | Process improvement | Note in task completion output for orchestrator |

Step 5: Report actions

Output in this format:

Compound: N actions
- 
- 
...

If no actions: Compound: no actions this session.

What NOT to do

• Do NOT run full /compound — that's for the orchestrator
• Do NOT manage K-entries in CORTEX.md — orchestrator concern
• Do NOT run consolidation or convention lifecycle checks
• Do NOT start new work — compound embeds learnings only
• Do NOT add >10 lines to any file without identifying lines to compress
• Do NOT make governance changes — note them for the orchestrator

Key files to update

• Project STATUS.md — if work changed project state
• Project TASKS.md — implied follow-up tasks
• Project ISSUES.md — workflow friction that reduced efficiency
• Task completion note — learnings and findings for callback

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: fangxm233
• Source: fangxm233/cortex-agent
• License: MIT

Install and usage instructions live in the source repository linked above.