L1 Improvement
- Reworked the skill into Skill Forge required sections with explicit triggers, stop conditions, and evidence logging.
- Added prompt-architect ceiling discipline and contract-style IO to keep iterations auditable.
STANDARD OPERATING PROCEDURE
Purpose
Manage recursive improvement cycles that collect failures, propose targeted changes, validate them, and stop when marginal gains level off.
Trigger Conditions
- Positive: iterative hardening of skills/prompts/agents, regression triage, A/B comparisons, or postmortem action items.
- Negative/reroute: net-new prompt design (prompt-architect/prompt-forge) or new skill scaffolding (skill-forge/skill-builder).
Guardrails
- Define stop criteria up front (delta threshold, timebox, risk acceptance) to avoid infinite loops.
- Capture evidence for each iteration: inputs, changes, tests, and results with confidence ceilings.
- Keep outputs in English; avoid hidden reasoning.
- Do not silently discard failed experiments—record them for future avoidance.
Execution Phases
- Intake: Identify artifact to improve, goals, constraints, baseline metrics, and stop criteria; classify constraints as HARD/SOFT/INFERRED.
- Hypothesis: Propose targeted changes addressing failures or goals; prioritize by impact.
- Apply & Test: Implement changes, run validation (happy/edge/adversarial), and log results with ceilings.
- Assess: Compare metrics vs baseline; decide to continue, pivot, or stop based on delta.
- Package: Summarize iterations, residual risks, and recommended next steps.
Pattern Recognition
- Quality drift → focus on regression tests and schema tightening.
- Safety issues → add guardrails, refusals, and escalation rules.
- Performance/latency complaints → simplify prompts and reduce tool calls.
Advanced Techniques
- Use multi-armed bandit style sampling for competing variants with limited budget.
- Apply self-consistency or debate to stress-test high-risk changes.
- Snapshot checkpoints so reverting is easy when metrics regress.
Common Anti-Patterns
- Iterating without baseline metrics.
- Changing multiple variables simultaneously, making results ambiguous.
- Ignoring ceiling discipline or failing to log evidence.
Practical Guidelines
- Limit each iteration to one or two focused hypotheses.
- Keep a changelog with timestamps, metrics, and confidence statements.
- Escalate to domain specialists when improvements stall.
Cross-Skill Coordination
- Upstream: prompt-architect/prompt-forge to clarify artifacts under test.
- Parallel: cognitive-lensing to unlock new hypotheses.
- Downstream: skill-forge/agent-creator to bake improvements into canonical docs.
MCP Requirements
- Optional memory/vector MCP to store iteration history; tag WHO=recursive-improvement-{session}, WHY=skill-execution.
Input/Output Contracts
inputs:
target: string # required artifact to improve
goals: list[string] # required goals/metrics
constraints: list[string] # optional constraints
stop_conditions: object # required thresholds/timebox
outputs:
iterations: list[object] # steps taken, tests, outcomes, ceilings
recommendation: summary # continue/stop with rationale
artifacts: list[file] # updated files if applicable
Recursive Improvement
- Meta: apply this SOP to itself; stop when improvement delta < 2% or risks documented.
Examples
- Harden a prompt that occasionally hallucinates by tightening schema and adding refusals.
- Improve a code-generation skill by adding edge-case tests and measuring deltas.
Troubleshooting
- No improvement after multiple cycles → revisit hypotheses or broaden search (new lenses/tools).
- Metrics regressing → revert to last good checkpoint and reassess constraints.
- Timebox exceeded → summarize current best variant and open risks.
Completion Verification
Confidence: 0.70 (ceiling: inference 0.70) - Recursive Improvement SOP rewritten with Skill Forge structure and prompt-architect ceilings.
1---2name: recursive-improvement3description: Drive iterative refinement loops with evidence capture, evaluation checkpoints, and stop conditions for skills, prompts, or agents.4---5
6
7
8
9### L1 Improvement
10- Reworked the skill into Skill Forge required sections with explicit triggers, stop conditions, and evidence logging.
11- Added prompt-architect ceiling discipline and contract-style IO to keep iterations auditable.
12
13## STANDARD OPERATING PROCEDURE
14
15### Purpose
16Manage recursive improvement cycles that collect failures, propose targeted changes, validate them, and stop when marginal gains level off.
17
18### Trigger Conditions
19- Positive: iterative hardening of skills/prompts/agents, regression triage, A/B comparisons, or postmortem action items.
20- Negative/reroute: net-new prompt design (prompt-architect/prompt-forge) or new skill scaffolding (skill-forge/skill-builder).
21
22### Guardrails
23- Define stop criteria up front (delta threshold, timebox, risk acceptance) to avoid infinite loops.
24- Capture evidence for each iteration: inputs, changes, tests, and results with confidence ceilings.
25- Keep outputs in English; avoid hidden reasoning.
26- Do not silently discard failed experiments—record them for future avoidance.
27
28### Execution Phases
291. **Intake**: Identify artifact to improve, goals, constraints, baseline metrics, and stop criteria; classify constraints as HARD/SOFT/INFERRED.
302. **Hypothesis**: Propose targeted changes addressing failures or goals; prioritize by impact.
313. **Apply & Test**: Implement changes, run validation (happy/edge/adversarial), and log results with ceilings.
324. **Assess**: Compare metrics vs baseline; decide to continue, pivot, or stop based on delta.
335. **Package**: Summarize iterations, residual risks, and recommended next steps.
34
35### Pattern Recognition
36- Quality drift → focus on regression tests and schema tightening.
37- Safety issues → add guardrails, refusals, and escalation rules.
38- Performance/latency complaints → simplify prompts and reduce tool calls.
39
40### Advanced Techniques
41- Use multi-armed bandit style sampling for competing variants with limited budget.
42- Apply self-consistency or debate to stress-test high-risk changes.
43- Snapshot checkpoints so reverting is easy when metrics regress.
44
45### Common Anti-Patterns
46- Iterating without baseline metrics.
47- Changing multiple variables simultaneously, making results ambiguous.
48- Ignoring ceiling discipline or failing to log evidence.
49
50### Practical Guidelines
51- Limit each iteration to one or two focused hypotheses.
52- Keep a changelog with timestamps, metrics, and confidence statements.
53- Escalate to domain specialists when improvements stall.
54
55### Cross-Skill Coordination
56- Upstream: prompt-architect/prompt-forge to clarify artifacts under test.
57- Parallel: cognitive-lensing to unlock new hypotheses.
58- Downstream: skill-forge/agent-creator to bake improvements into canonical docs.
59
60### MCP Requirements
61- Optional memory/vector MCP to store iteration history; tag WHO=recursive-improvement-{session}, WHY=skill-execution.
62
63### Input/Output Contracts
64```yaml
65inputs:
66 target: string # required artifact to improve
67 goals: list[string] # required goals/metrics
68 constraints: list[string] # optional constraints
69 stop_conditions: object # required thresholds/timebox
70outputs:
71 iterations: list[object] # steps taken, tests, outcomes, ceilings
72 recommendation: summary # continue/stop with rationale
73 artifacts: list[file] # updated files if applicable
74```
75
76### Recursive Improvement
77- Meta: apply this SOP to itself; stop when improvement delta < 2% or risks documented.
78
79### Examples
80- Harden a prompt that occasionally hallucinates by tightening schema and adding refusals.
81- Improve a code-generation skill by adding edge-case tests and measuring deltas.
82
83### Troubleshooting
84- No improvement after multiple cycles → revisit hypotheses or broaden search (new lenses/tools).
85- Metrics regressing → revert to last good checkpoint and reassess constraints.
86- Timebox exceeded → summarize current best variant and open risks.
87
88### Completion Verification
89- [ ] Stop conditions defined and respected.
90- [ ] Iteration logs include tests, results, and ceilings.
91- [ ] Recommendation provided with residual risks.
92- [ ] Artifacts updated or explicitly unchanged.
93
94Confidence: 0.70 (ceiling: inference 0.70) - Recursive Improvement SOP rewritten with Skill Forge structure and prompt-architect ceilings.
majiayu000