We all know the pain of running coding agents on large projects. They start off strong, but as the context window fills up, they eventually forget previous work and start breaking things. I just read a brilliant post from a savvy developer on Reddit who solved this with a method called "searchable anchors." It keeps agents on track without complex tools.
The Key Idea
The core concept is to force agents to leave a trail of breadcrumbs for their future selves. Instead of relying on their fading short-term memory, the creator sets up a rigid protocol where every agent documents its work in a /docs/ folder using specific "anchors." Before any new agent starts a task, it doesn’t just guess: it consults a master index and searches these anchors to find the exact context it needs. It turns basic text search into a perfect, infinite memory system.
📌 The "Anchor Protocol" Rules
The author shared specific instructions to add to your system prompt to make this work. You simply tell the agent:
"Before starting, read /docs/ANCHOR_MANIFEST.md, grep for anchors related to your task, and read the matching files."
When the task is done, the agent must create a new markdown file summarizing its work using a unique tag format, like <!-- anchor: feature-name -->. This strict input/output cycle ensures that Agent 50 knows exactly what Agent 1 did.
💡 Documentation That Actually Works
Most AI summaries are either too vague or way too wordy. The expert points out that for this to work, the doc files need to be informative but concise. They should list every file path touched, specific function names, and key implementation decisions. It’s not a diary; it’s a technical map. The rule is that any agent reading it should understand what exists and how it connects in under 30 seconds.
✅ Scaling Without Confusion
The beauty of this system is that it bypasses context compaction, where details get crushed and lost over time. Because the agent searches for fresh data every single time, the memory never degrades. The original poster notes that this method prevents the dreaded hallucinated file path issue, giving Agent 200 the same clarity and accuracy as Agent 1!
This logic is simple but incredibly effective. I highly recommend checking out the full source to copy the exact prompt template.
💡 FAQ & Troubleshooting
Why do coding agents start hallucinating on large projects?
As the context window fills up during long sessions, memory compaction occurs. Agents lose track of previous actions or work completed by other agents. This leads to them hallucinating file paths that don’t exist or overwriting code because they cannot recall the current state of the project. The Anchor Protocol solves this by forcing agents to read a persistent “map” (manifest) before acting.
What is the strict format for creating searchable anchors?
Anchors should use HTML comment syntax to remain invisible in rendered Markdown but visible to grep. The required format is <!-- anchor: feature-area-specific-thing -->. To ensure reliability, the anchor text must be lowercase, hyphenated (no spaces), under five words, and unique across the entire project.
What specific information must be included in the documentation files?
Documentation files should not be generic summaries or verbose diaries. To be effective, they must include:
1. File Paths: Every file touched by the feature.
2. Key Elements: Important function or class names.
3. Connections: How files interact (e.g., “authGuard.js calls jwt.js”).
4. Decisions: Key implementation choices (e.g., “chose cookies over localStorage”).
The goal is for an agent to understand exactly what exists and where it lives without reading the code.
Can I use structured formats other than Markdown for system definitions?
Yes. While Markdown is standard for the narrative documentation, structured formats like TOML, .cfg, or C-style syntax are excellent for defining system metadata, service layers, and core rules. For example, TOML allows you to clearly define layers (e.g., [stack.layers]) and operational rules in a format that ensures strict adherence by the agent.
I’ve discovered ‘searchable anchors’ in prompts, coding agents cheat code
byu/MrCheeta in