Skip to content

SKILL: Trace Capture

Skill ID: skill-trace-capture Canonical location: AAA/wiki/skills/SKILL_TRACE.md Status: CANONICAL — this is a meta-skill, not a draft When to use: After EVERY non-trivial task — before session end or before moving to new work Severity: LOW risk_band — this skill only writes evidence, never modifies canonical pages

Summary

After completing any non-trivial task, emit four artifacts: 1. A trace/event log of what happened 2. An evidence bundle of raw outputs and receipts 3. A proposed wiki delta (what should change in TREE777) 4. A confidence/risk tag (LOW / MEDIUM / HIGH / 888_HOLD)

This is the entry point to the promotion ladder. If you only do one thing after work, do this.

Trigger Conditions

Always run after: - Any task that took more than 5 tool calls - Any task where the agent made a decision (not just followed a script) - Any task that produced a non-trivial output (file creation, API call, config change) - Any task where the agent was uncertain and made a judgment call - Session end (end of any non-trivial session)

Can skip if: - Task was a simple read-only query (just looked something up) - Task was trivially simple (one tool call, obvious outcome) - No new information was produced

Procedure

Step 1: Compose the Trace

Write a plain-text trace of what happened. Format:

TRACE: [YYYY-MM-DD HH:MM] | AGENT: {agent-name} | TASK: {one-line task description}
---
Step 1: [what was attempted]
Step 2: [what happened]
Step 3: [outcome]
---
Confidence: [LOW / MEDIUM / HIGH]
Risk: [LOW / MEDIUM / HIGH / 888_HOLD]

Example: 

TRACE: 2026-05-17 14:30 | AGENT: claude-code | TASK: Patch Gemini config with SPATIAL_LAW
---
Step 1: Read Gemini config location (~/.gemini/SYSTEM_MD.md)
Step 2: Appended SPATIAL_LAW block to end of file
Step 3: Verified with grep — match found
---
Confidence: HIGH
Risk: LOW

Step 2: Collect Evidence Bundle

For each significant action, capture: - The exact command or API call made - The raw output (not summarized, not interpreted) - The verification result

Store in wiki/raw/notes/{YYYY-MM-DD}_{agent}_{task-slug}.md:

# Evidence: {task description}
**Agent:** {agent}
**Timestamp:** {ISO 8601}
**Task:** {what was attempted}

## Raw outputs

### Command 1: {description}
\`\`\`
{raw output verbatim}
\`\`\`

### Verification: {what was checked}
\`\`\`
{verification output verbatim}
\`\`\`

Rule: Capture raw outputs, not interpretations. Let future readers interpret.

Step 3: Write the Proposed Delta

Based on what happened, what should change in TREE777?

Situation Proposed delta
New failure pattern observed wiki/scars/scar-{incident}-{date}.md draft (status: proposed)
New successful technique discovered wiki/skills/skill-{name}.md draft (status: proposed)
Existing skill out of date Update wiki/LOG_MD.md with correction proposal
New concept needed wiki/concepts/concept-{name}.md draft (status: proposed)
Evidence that existing page is wrong Write correction to wiki/raw/notes/ + propose via log
No change needed State "no TREE777 delta"

Format for proposed delta in log: 

## [YYYY-MM-DD] PROPOSED | {agent} | {delta description}
- Page: {which page}
- Change: {what should change}
- Evidence: {where raw evidence is stored}
- Confidence: {your confidence this should change}
- Risk: {impact if wrong}

Step 4: Assign Confidence/Risk Tag

Tag When
LOW Task was routine, outcome was expected, no surprises
MEDIUM Task had complexity, some uncertainty, outcome was reached but not trivially
HIGH Task had novel elements, significant uncertainty, or non-obvious outcome
888_HOLD Task revealed something that should not be changed without 888 deliberation — escalate immediately

Step 5: Append to wiki/LOG_MD.md

# Format for log entry
## [YYYY-MM-DD HH:MM] trace | {agent} | {task summary} | {outcome}
- Confidence: {tag}
- Risk: {tag}
- Evidence: {path to evidence bundle}
- Proposed delta: {description or "none"}

Preconditions

  • Task completed (at least attempted)
  • Terminal/file access to write to wiki/LOG_MD.md and wiki/raw/notes/
  • Basic awareness of what "non-trivial" means (5+ tool calls, decisions made, novel elements)

Expected Outputs

  • wiki/LOG_MD.md entry appended with trace summary
  • wiki/raw/notes/{YYYY-MM-DD}_{agent}_{task-slug}.md evidence file created
  • Log entry includes: confidence tag, risk tag, evidence path, proposed delta (or "none")

Failure Modes

Failure Mitigation
No significant output to capture Still log the attempt — "task attempted, no novel output"
Can't write to wiki (permission) Write to local tmp, propose in session summary
Evidence too large to store Summarize key signals; note full output available in session logs
Unsure whether task was "non-trivial" Default to capturing trace — safer than missing evidence

Relationship to Other Meta-Skills

This skill is the entry point to the promotion ladder:

skill-trace-capture (this skill)
    ↓ emits evidence bundle
skill-scar-distill (next skill)
    ↓ if failure pattern detected
skill-skill-promote
    ↓ if pattern repeats 2-3x
→ 888 JUDGE promotion
  • [[TREE777]] — governance framework this skill implements
  • [[skill-scar-distill]] — next skill in the ladder (converts failures to scar drafts)
  • [[skill-skill-promote]] — skill for promoting patterns to canonical
  • [[concept-memory-knowledge-paradox]] — the bidirectional mirror this skill feeds

DITEMPA BUKAN DIBERI — Capture first. Prove later. Always.