Forced Feedback Loop

English | 한국어 | 日本語 | 中文

An agent skill that forces continuous self-feedback loops until the budget (time or loop count) is exhausted, steadily improving work quality without ever stopping voluntarily.

Problems This Solves

• AI does one shallow pass and declares "done"
• Hollow offers like "I can also do X next" instead of actually doing it
• Work history lives only in ephemeral context and vanishes after the session
• Confirmation bias — the agent validates its own conclusions without real testing
• Search space narrows over time when the agent only recombines existing information
• Knowledge files grow monolithic and become unsearchable as ideas accumulate

Core Principles

1. Never stop before the budget is spent — active work only during min_required_minutes and/or min_required_loops
2. Generate new proposals every loop — both convergent (recombine known) and divergent (explore unknown)
3. Write everything to files — reverse-chronological reports in work-log.md, session continuity guaranteed
4. Manage knowledge as individual notes — Obsidian-style KB with per-idea files, searched by a dedicated sub-agent

Skill File Structure

forced-feedback-loop/
├── SKILL.md                        # Core execution policy (loaded on activation)
└── references/
    ├── DOMAIN-TABLES.md            # Domain-specific tables (lazy loaded)
    ├── PROHIBITIONS.md             # 27 prohibited behaviors (lazy loaded)
    └── KNOWLEDGE-BASE.md           # KB system: note format, tags, search protocol

Runtime Artifacts

The agent creates these files in the working directory during execution:

<working-directory>/
├── work-log.md                     # Full work history (reverse-chronological report stack)
└── kb/                             # Knowledge Base (Obsidian-like)
    ├── _index.md                   # Tag → file auto-index
    ├── raw/                        # Unverified ideas, drafts
    │   ├── 20260303-143000-xxx.md
    │   └── ...
    ├── archive/                    # Rejected items (with reason)
    │   ├── 20260303-150000-yyy.md
    │   └── ...
    └── curated/                    # Validated items
        ├── 20260303-160000-zzz.md
        ├── UNCERTAINTY-REGISTER.md # Uncertainty register
        └── GAP-REGISTER.md         # Data gap register

Why Individual Notes?

A monolithic 3-file approach grows unwieldy as ideas accumulate:

• Loading everything into context wastes tokens
• No way to retrieve only relevant items
• Search efficiency degrades as sessions get longer

The KB system stores each idea/decision/evidence as an individual Markdown file with YAML frontmatter (tags, status, confidence, etc.), so the search sub-agent can retrieve exactly what's needed.

Search Sub-Agent

When the agent looks up information in the KB, it delegates to a search sub-agent instead of reading everything:

1. Tag-based lookup via _index.md
2. Read matching note files (max 10)
3. Return top 5 results as {id, title, status, confidence, summary}
4. Include relevant evidence, contradictions, and gaps

If the KB has fewer than 10 notes, the agent reads files directly without a sub-agent.

Supported Domains

task_type	Use case
research	Investigation, literature review
code	Development, engineering
design	UI/UX design
document	Writing, documentation
analysis	Data analysis
project	Project management
other	Anything else

Evidence types, source hierarchy, tag systems, and test methods adapt automatically per domain. See references/DOMAIN-TABLES.md for detailed mappings.

Usage

Installation

Extract the downloaded zip and place it in your agent's skill directory.

Agent	Path
Perplexity	Upload via User Settings
OpenAI Codex	~/.codex/skills/forced-feedback-loop/
Claude Code	~/.claude/skills/forced-feedback-loop/
Augment	~/.augment/skills/forced-feedback-loop/
VS Code (Copilot)	.agents/skills/forced-feedback-loop/ (project root)

Examples

Research:

Spend 30 minutes researching "EV battery market trends in South Korea".

Code:

Spend 20 minutes improving this API server's response time. Current p99 latency is 500ms.

Design:

Spend 15 minutes improving the usability of this dashboard layout.

Document:

Spend 25 minutes strengthening the argument structure of this technical proposal.

Count-based (no time limit):

Run 10 feedback loops to improve this function's error handling.

Both (time + count):

Spend at least 20 minutes AND run at least 5 loops optimizing this query.

Specify a time budget, a loop count, or both. The agent activates the skill automatically and runs feedback loops until the termination condition is met.

Termination Modes

min_required_minutes	min_required_loops	Behavior
set	—	Time-based: run until time expires
—	set	Count-based: run exactly N loops
set	set	Both: run until both thresholds are met

If neither is specified, defaults to 5 minutes.

Convergent × Divergent Dual Proposals

Each loop's proposal step alternates between two modes:

Mode	Direction	Description
Convergent	Known → new hypothesis	Combine 2+ existing findings from the KB into a new proposal
Divergent	Unknown → new hypothesis	Identify an adjacent question the KB can't answer yet, then actively search for new information to test it

At least 1 divergent proposal must appear in every 3 consecutive loops. Pure convergence narrows the search space. Divergent proposals must pass a relevance check against task_goal, so they stay on-topic.

Reading work-log.md

Reports stack newest-on-top. The first line of each report declares its line count, enabling incremental reading:

=== Report #3 | lines: 6 | elapsed: 14:30 | type: feedback ===   ← line 1
DIAGNOSE: Weakest decision is D2 (confidence 0.4).
PROPOSE: Inversion — what if we use push instead of pull?
TEST: grep codebase for event-driven patterns → 3 hits.
UPDATE: D2 confidence raised to 0.65.
META: Progressing. Next loop targets D4.                         ← line 6
---
=== Report #2 | lines: 4 | elapsed: 10:00 | type: feedback ===
...

lines: rule: Count from the header line (inclusive) to the last content line (inclusive). The --- separator and blank lines between reports are excluded. The agent verifies this value after writing each report and runs a lint check in scripting environments.

Cross-references use relative coordinates (K-reports-below, line N below), so references survive log compaction.

Multi-Session Continuity

If work-log.md and the kb/ directory remain from a previous session, the agent picks up where it left off. It reads kb/_index.md, uses the search sub-agent to review recent notes, restores context via incremental reading, and continues the feedback loop.

Continue the battery market research from yesterday for another 20 minutes.

Compatibility

Follows the Agent Skills open spec and works with any agent that supports this spec.